Analyze Product Data with Claude Code Opus 4.6

Speaker: like, drop something in the chat like. Like, besides yes, like, I don’t know, drop, like, where you’re… located or something, or… I don’t know what’s the weirdest thing on your desk right now? is kind of fun. I have a walkie talkie on my desk. It’s a radio thing for snowboarding. I have a lot of stuff on my desk, but it’s kind of a mess. Let’s do where you’re… where you’re from, what do you do? And then, I guess, once on your desk? Well, I like his banana chips. Oh, we got Munich. We got Texas. Wow. aligners, we got Dubai, Singapore, Toronto, San Francisco. Barcelona, dang. Whoa. It’s like, hey, Attendee, what’s up? There’s a lot of cool people in here. Northern Utah.

Where in Northern Utah, Attendee? India, we’re in India? We’ve got side. We know where you’re at. North Logan. Nice. I’m… I’m in Tahoe, South Lake Tahoe, California. Love it. Surabi and higher in the bay. Hi from Madrid. Man, you guys are up late. You guys really like analyzing data with Claude code. so international. Uh, cool. People are coming in all right. It seems like people can see my screen and hear me. So. We can jump right into it. Okay. First, I want to start with a little. This is going to be a different lightning lesson than I think is usual. I think it’s. We’ll get more into this in a bit in a couple slides like originally we had this planned for later in March.

Hey, Attendee from Louisiana. And we… pump this up a lot earlier, like a month earlier after we did a bunch of testing with Opus 4.6. So basically a few weeks ago I took me highest Ravi all took time off work off our day jobs to be like, Hey, we gotta. Kind of, like, see what’s going on, and we got in person, I went down to San Francisco, and we tested out a bunch of the model capability, and it was, like. way beyond where it was previously. And so we pumped up this. This, uh… this lightening lesson. And so, like, we’re gonna go through some demos of, like, what we’ve used so far. This isn’t… We’re not selling you a product.

Like, everything we are going to show you, we’ve open sourced it for free. We’ll drop a link in there. I don’t know if higher you can even drop that in now. I think probably… like, the reason it’s free is because I think anyone here can build it. Anyone can build it to the context of their own, and I think it’s gonna get more and more powerful. I do want to have like a couple, like. things I know probably a lot of people saw the news recently with block. laying off half of their workforce, and a lot of that was because they’re. saying they can leverage AI to do those people’s jobs. That’s not what this is about.

So if you’re if you’re here to figure out how you can lay people off and have AI do the job, maybe like leave. That’s not what this is. So what we’re trying to do here is teach you. how to automate as much as your job as possible as data professionals to get yourself extreme agency so you can go do a more interesting job and kind of get ahead of all this. Um… We don’t want this to, like, replace people like we want it to replace some of the tasks that we have done previously. It’s gonna be it’s like a little bit of scary time. I think the way I think about it is like the some of the tasks that comprise like my job over the past decade do not define.

like, my identity or my professional identity, or who I am. And so, like, there’s a very new, interesting job. I think that’s coming, but I do think it’s a critical that we like lean into this technology and learn how to use it. I think there’s going to be people who are like, just like SQL and Python, like beginner mode, and people who are like. Expert mode and, like… There’s still gonna be a world for data scientists and smart people. We’re just going to be leveraging different tools. But I do think it’s critical that we learn how to leverage them and experiment with them drastically, because if not. Someone else is gonna do it. Someone high up is gonna.

see how to automate out your job, and it’s gonna get left behind. So it’s like, I don’t know, I am, like, uh, both… optimistic doomer to some degree, I guess. But all right. That’s my preamble. Let’s get into it. So thanks for being here. I’m Shane. I’m joined today by my colleagues, Hai Guan, head of data at ENTRE, Sravya Matapali. She’s data leader at Superhuman. Do you want to introduce yourself real quick? Sure, uh, I can go ahead. Hello, everyone. I’m Sravio Maripali. I lead. I’m a senior data science manager at Superhuman. Previously it was called Grammarly. I’ve basically started my career in data science at Microsoft. I was there for like 6 years.

Yeah, and later worked at companies like eBay, Nextdoor. That’s where Shane Hai and I met. And we decided we want to do something together to help the data ecosystem. And here we are. We have a podcast where we run, talk with a bunch of data science leaders and, like, overall data leaders. And here we are. We got ourselves into understanding what AI could do with data, and we had to share it with the world. So glad we have all of you here today to go through what we’re going to share. Yeah, you can go. Hi. Hey everyone, my name is Hi. I’m the head of data at a legal tech startup. Uh, similar to Shane at ENTRE.

Uh, I was previously leading data science teams at, um, uh, social media companies like Nextdoor, LinkedIn, Pinterest, Meta, and so, uh, been seeing a lot of changes, certainly, and this is now… a very, uh, exciting times, and we really are… wanting to sort of, like, bring the capability to all of you to make sure… the ones that are sort of like, you know, getting ahead of the curve. will have the best, you know, like, you’re gonna really, like, 10x your career, so that’s what we’re here for. Cool. Thanks, guys. So we’re going to be dropping a bunch of links throughout this discussion, and I’m sure a lot of people will have questions, and we’ll be dropping the same links again and again.

A lot of that stuff. If you go to AI analyst lab.ai. Maybe you could drop that in the chat, Haya Sravya. It’s on the slide here. Yes, I already did. It’s yeah. A lot of nice, nice. Okay. So like it’s a lot of this stuff. That’s our website. A lot of the stuff’s linked out from there. So our whole thing is like try open source, make as stuff free as possible. And then for people who want like hands-on training. We do a boot camp next month. That’s going to be a weekend boot camp, 2 days live. No async stuff like all live to teach you how to build your own system of what we’re going to do today. And then we have a longer course in April. Around… one around AI like.

Hey, analytics for builders was more like analytical frameworks and thinking as well as a bit of cloud code. So it’s more like the full end-to-end, like, how do you leverage AI in the data science or data professional? um, cycle. Check those out. Our our ethos is kind of like anything we build anything we make any information we have, we give it out for free, and then, if it’s something that’s like, I gotta take hours out of my day. I mean, I don’t know. This is free, I guess. Then, like, obviously, we charge for that right? So that’s why this course is there. So check that out. Ai analyst.ai. and let’s get into it. So. Quick poll to start. Maybe drop in the chat. What percentage?

Are the analysis questions your team gets? Do you think actually get answered? Like not just acknowledged, but answered with validated numbers and a recommendation. And this could be the percent that’s coming from your stakeholders. the percent of ideas that you just have, and your team, like, all these things you want to work for, all these questions you have, um, within, like, a year. Oh, there’s some variance in here. So we’ve got some people who are 80%, some are 20%, some are 60. Man, 80% is nice. Uh, 30%. Yeah, any… anything that comes 60 to 70%, 50. I’m not seeing any hundreds yet. Okay, so a hundred percent is not there. 60 to 70%. Attendee. Okay. yeah.

So if your number is under 100 like 80% is pretty good. But if your number is under 100, like. Obviously, you can see here, you’re not alone. If your number’s under 50, you’re not alone. There’s probably a lot of questions that people aren’t even asking, right? Because they know you don’t have time to get to them, because you can’t even get to the 100%. So we’ve all been here. Someone asks like. like someone asks like, why did retention drop? And maybe you spend like a few days writing SQL, checking edge cases, building charts, and you deliver it, and then they ask the same question cut by segment. or for another day, or… and then by channel, and then another day.

And that’s, like, the questions that get answered, right? The ones that don’t, they pile up in backlogs or in our minds. They get… they get answered with a gut feel instead of data, and it’s not because anyone’s lazy, right? It’s it’s because. The execution cost is too high for the volume of questions that come in either internally or externally. Um… So here’s the other piece. When you do answer them, it’s like, how much time. How much time was spent thinking versus typing? So for most folks, at least for me, like, I would say it’s somewhere around like 80% mechanics writing queries.

uh, checking joins, validating those queries, formatting charge, and then 20% is actual thinking, interpreting the results, deciding what matters, figuring out what to do about it. And that’s like the most important part and also the most interesting part. And that’s where I want the entire field of the data profession to. totally go. So we want to flip that ratio basically by leveraging AI. So we want to flip. 80% mechanics, 20% thinking to 80% thinking 20% mechanics. Same hours. More interesting work. And I think a lot of companies are not going to do this.

I think a lot of companies are really going to screw this up, and they’re going to stay at 80% mechanics and 20% thinking, except they’re going to make everyone 5x mechanic they’re mechanics, and 5x their thinking. Or they’re going to have a smaller workforce and they’re going to like lay off a bunch of people and just still have people do, like, again, 5x. the the the mechanics, 20% the thing. I think that’s the wrong way to do it. I think we need to flip it. And then, once we get to 20% mechanics, and we have 80% thinking, we have all this free time we earned like, I think time is all you need for thinking, and… We can discover, like, what’s the next job?

We can go solve the problems end to end the recommendations ourselves that are coming out of these analyses. We can start thinking about. The questions and the problems that came up that could never come up before because there’s the requirement is that all these prerequisite questions that we could never solve because we didn’t have enough time are now solved. So that’s kind of like like my whole thing is like, I want to get data professionals like extreme agency, so they can go far beyond their kind of like. mechanic’s job to like like same hours, but way more interesting work. And so we actually, we talked to our friend, like.

Shane Cook, a former colleague of ours the other day, and he had a really good insight where it’s like, in a world where execution and building is extremely fast and extremely cheap, and anyone can do it, and you no longer have this moat of like. Oh, our engineers already did people like are really fast on mechanics. Then, like the most important thing is ideas. The most important thing becomes thinking, and that’s the people who I think will win out, and you have to have time to do that. dedicated blocks of time. So that’s kind of what we’re aiming for here as we automate things. I don’t know if that resonates with anyone, let me know in the chat, or it’s you think I’m crazy, it’s fine too.

So. We’ve been building analytical systems like high Sravi and I and Claude Code since last year. And honestly, for a lot of that time it was a lot of hand holding. So you’d ask a question. Uh, the system would try to frame it. You’d have to correct that framing. It’d write some sequel, you check the SQL. It was pretty good at SQL, actually, like framing questions and knowing, like, what to ask and forming hypotheses, how to lean in a lot, but, like, code stuff like SQL. You could get that pretty good with, like, semantic views and stuff, and just setting up a nice found data foundations.

It build a chart like still had to do a lot of back and forth with that like you have to fix the chart a lot. Every step, though, like, required, really, like, someone in the loop end-to-end. It still saved time, a lot of time. This is like basically what our we’re going to teach a lot in our course. question that took you like you know 3 days might take you one day. But you were there for most of those hours directing, correcting, validating. It was less like having a. the analyst, and more like having a really fast intern who needed constant supervision. We are originally… plan this lesson for late March, as I mentioned, but we moved it up because like that hand holding.

isn’t as much there anymore. So high like 30 seconds. This is for anyone who hasn’t used it. We’ll we’ll kind of like talk about what Claude code is. Actually, let’s do this in the chat. Okay, we’re going to skip the question of like who hasn’t heard of cloud code? Because cloud codes in the title of this session. So hopefully you have heard of it before, but drop a one in there. If you have heard of cloud code, but you haven’t used it yourself. drop a 2. If you’re using Claude code in your workflows, but you’re not using it to analyze data. In a 3, you’re using Claude code and you’re using it for data analysis. lots of ones, twos. Oh, a few threes. Nice. Yeah, a lot of people are.

There’s some people in here, you know, they okay? So it’ll be interesting to see how this resonates with you, or are you using it for data. Cool. Uh, let’s keep going. So basically, I’m not talking about like Claude in your browser, like chat stuff. Right? I’m talking about… this is different. So, like, that one, I feel like in chat the way to think about it is like Claude has a brain. And then Claude code. Uh, I feel like it has hands now. So it can read files on your machine. It can write Python, run it. It can see the output and decide what to do next. That is that is the critical part that, um.

I think we learned this past couple of weeks is like sometimes when we decide what to do next, it wasn’t very good. And I was like, Hmm, quite quite good at deciding what to do next if you nudge it and give it some direction. So it’s executing code in a real environment with real data. And the piece that matters most for what I’m about to show you is the Claude MD file. So that’s a file you put in your project and it tells Claude who it is, what it can do. Uh, what your data looks like.

It’s… think of it as like an analyst onboarding doc, like when someone comes into their new job, and it could be a junior analyst, it could be a senior staff principal or manager like even you could be the best analyst or data scientist in the world. You come to a new job, and you don’t know the context of the business, or the people. Or the data, like, you’re not much use. You need to onboard. So cloud defile, it’s like it’s onboarding doc, but it doesn’t just get this onboarding when it first joins, it reads that file at the start of every session. So we’ll come back to this later. So what changed 2 weeks ago? So that’s Opus 4.6. This is.

the new model by anthropic came out and basically like it can kind of hold that entire analytical workflow in its head instead of having to. Do you one step at a time. Okay, yeah, Attendee, reason I use Claude code, and I don’t know if you can do this in codex yet. I haven’t screwed around with it that much. I think I heard that the 5.3 is actually, like. on par with Opus 4.6. I like Cloud Code because you can spin up a bunch of sub-agents, and so you can be running, like, tons of agents in parallel or sequentially. But to be fair, I’ve just been, like, super focused on that the past two weeks that I haven’t looked elsewhere. Um… Does that answer your question? Um, cool.

So yeah, not one step at a time. It can carry that whole framework. like, from end to end, I think. So, you know, it can frame the question, it can decide what data it needs. Frame the question is very interesting. This is a skill that I was convinced data scientists would be needed to frame the question. You know, when you get, like, very obscure questions. That’s like. How’s our, like, engagement doing? Like, you can now make agents and skills to be like. I’ll show you later in the MDE files, but, like, to be like, okay, here’s their question. What do you think they really need? How would you make that actionable? Like, if it was actionable, what are hypotheses you would create?

Given the actionability and the hypotheses, frame new quest… business questions. And it’s like. Dang, I have I spent so many years learning those skills, and now I can just write them in a markdown file and Claude can do it. But I think we’re still needed there a bit because it can go down off some silly business questions too. But I can frame the question. I can decide what data it needs. It can write the SQL. I could run the SQL, it can check whether the output even makes sense. It can catch its own mistakes, it can build the charts, it can write the narrative, and it can tie it all together. Not out of the box, not if you just open Cloud Code, well, does it do all this stuff?

You have to make a bunch of like. agents and skills and and Python helpers, and by you make it, I mean, work with Claude to make it. You don’t have to type anything. Um… But you do have to, like, teach it and grow it just like you would teach and grow an actual analyst. Um, and it doesn’t do it perfectly. I want to be clear about that. It makes mistakes. It really needs you like the nature of this, but the nature of the mistakes can now be changed, so… the way I kind of think about it is before it felt like it was always getting lost when we would try these over the past 6 months or so, and now I feel like it occasionally will like stray off the wrong path like. I don’t know.

I do a lot of trail running, and I also do these like. weird 100 mile races where I’m like trip running in the middle of the night and. uh, you know, never have I got into a situation where I’m like, dude, I don’t know where I am right now. I might do that if I didn’t have my GPS watch, though. I might get myself into a situation where, like, dang, I strayed down some stuff, and another path and another path, and I just don’t even know where I am. This is kind of scary. I don’t have service. Uh, now I will definitely go down paths, but I have, like, my map on my Garmin, and it’ll ping me after like a quarter mile and be like, Hey, you’re off course.

And it kind of sucks when you’re at, like, mile 60 and you’re like, dang, I just want an extra quarter mile, I have to go back. But I don’t get lost, right? I just have to recalibrate, and that’s what I think is happening a lot more with this model than the previous. You can really recalibrate them. So it occasionally will pick the wrong column, or misinterpret a business rule, but those are correctable. And those are the kinds of mistakes like anyone coming into a new company and analyzing their data would make. And like the first couple months or a few weeks on a data set.

And the really cool thing with this is, like, documentation is so cheap and easy to just have it write those corrections. into its, like, Cloud MD file or a knowledge base, so it doesn’t make that mistake again, or makes it, at least, has a much lower probability of making it again. So the bottleneck, I think, has shifted. Now from… execution to judgment. That’s like the the flip. So let me let me kind of show you here. We’ll go over this example in a minute. So we have a synthetic data set we’re going to use. I’m not going to Claude code and. It’s called, like, Novomart. It’s like a… synthetic rip off of Amazon, basically.

So it’s e-commerce data has like 50k users, 13 tables, and a full year of behavioral data. So stuff like transactions, sessions. marketing attribution, customer segments, like, the works. And so we can ask it very challenging questions, very simple questions. The time it takes to go through this and the workflows it triggers and the depend on the the challenge, the level and depth of the analysis. So this one would be. pretty difficult one. Something like, why did conversion rate drop 4 points between Q1 and Q2? And which customer segments drove that decline. So it’s not really a softball, right? It’s like it requires joining multiple tables.

It’ll require calculating rates across time periods, segmenting the population. Uh, figuring out which segments explain the aggregate change like a solid analyst could easily spend a week or two on doing a deep dive around this question. Um… Especially if they’re being thorough with validation. And so it’ll take Claude a while to do this, too. So like this might be an analysis that takes quite an hour to do. But like 2 weeks to an hour is pretty good. We’re gonna start with some softball ones for it, and I can show you the output of this later. But like simple things will be. pretty fast. So let me switch screens here, and I shall show you. Claude. Okay. Can everyone see this screen? Yes.

Yes. Okay. So. I’m in VS Code right now, but you know, you can use just your terminal directly. You could use any IDE you want, right? You could use cursor or anti-gravity or. You know, I don’t know. There’s a bunch of them, right? You could put this. This is this is local right now, but you could be in SageMaker or some hosted environment. Um, and so to get in here, like, let me just show you actually. So this is just a terminal. and all you have to do is type like bod. and it’s going to open up Claude for me. Yeah, I trust this blah, blah, blah. Um… Let’s do a few things first. So.

The first thing a mind shift I think you have to have when you’re working with this stuff is like you just ask Claude stuff about what’s going on here. So we’ll get into this later. But I might ask something like. Hey, hey. I’m sharing… the AI analyst with a bunch of people right now. I want to show them how this repo works, how the AI analyst works. Can you write me a TLDR? And also make me a workflow of how it works in ASCII art. I need to paste that in. Okay, so I just used this thing called Whisper flow. So it’s like. speech to text, so you can, I’d say it’s like 3 or 4 times faster than typing, and it’s nice, because then you can, like.

just talk to it directly a lot easier than having to type these things out. So it’s going to stop me along the way and be like, are you? Am I allowed to like, do stuff? It’s just not going to go off the rails and start. looking at things. Um, while this kicks off, this is the cool thing. You can just go to other terminals, right? So I have this other terminal open. And let’s kick off some analysis while it’s figuring out how to teach us about what it is. So… First thing I like to do. And once you have your stuff set up, you don’t necessarily have to do this, but I like to kind of prime it to make sure it’s looking at the right data. So I’m going to say, like.

Hey, tell me about the data sets available here, like the Novomart data set. And so it’s going to kick off and start learning about this. So it’s. It’s looking up available stuff in my working directory. It’s reading a file, reading another file. I like if you hit control O, you can kind of see what it’s actually running. Right. So it’s thinking, let me look at the knowledge about data sets and have our data set specifically. It’s pretty nice. You can see what’s thinking and how it’s reasoning through things. But if it’s running like SQL, it’ll like also start showing you like the queries it’s actually running. So you can kind of spot check.

along the way, see where it’s going, and you can always, like, hit Ctrl-C and stop it. If it’s going some weird direction. Alright, cool. So it’s got a bunch of stuff here for us. So currently connected to Novomart. It’s a synthetic e-commerce. I already told you guys about all this stuff. Yeah, TLDR. It’s too long. Don’t read. Um, so it tells us all the tables, has users, orders, order items, products, sessions, events, etc, etc. key dimensions for analysis, device acquisition channel product category. Good to know the events table is large. So this is kind of cool, right? So like, it’s also knows like some gotchas and things, how to work with it like events tables large.

Use a sessions table when when possible for performance. Some columns are stored as float due to nulls cast them as int before joining. There’s also example data sets and examples if you want to explore beyond Novomart. And so some of the reasons it knows how to like. Do these double checks on, like, the data stuff. That’s not necessarily going to happen out of the box. It’s because you’re going to have to write or work with it to create an agent or a skill to know that when I ask you to go look at a data set like you go look for all these like. potential errors and caveats that we need to be thinking about. We have to think about… I want us to figure out all that stuff.

Before we do any analysis, because that’s the way I think about it, you want it to work how you would work and or how an analysts would work before any like what’s going to make my analysis. better than someone who has the same experience as me and the same skills as me, but I’ve been in a company for a year, and they’re on week two, is that I just know all the screwed up stuff about our data and things like that. So we’re trying to teach it to do that. Okay. We’ll come back to this, but let’s go back to our other terminal. Because I want to tell you what it does. So… Here’s where we were asking about how this works.

So Tldr, the AI analyst turns a plain English business question into a validated slide deck in minutes. It’s exaggerating on in minutes. It’s very okay, like maybe a very simple question. 15 min for a simple deck. If you do something pretty hard, it’ll take, like, an hour or so. But that’s, like, without optimizing it. You ask something like, why did conversion drop? And it runs four phase pipeline frame your question, analyze story and deck. Uh, it self-validates every step as much as it can. You still need to do a validation. It learns from corrections, it works with your own data, and ships as open source. and then I love doing the askii art thing. This is like a recommendation.

I would say, anytime you’re building something in Claude like I just find this super helpful for it to tell me how it, how it goes through stuff. So. Right here, right? It says frame and live story deck, and it tells me, like, what it’s going to do with each of those. If I wanted to go further, I could be like, we’ll come back to this, how this. If I was to run the full end-to-end pipeline. Could you make an ascii art of what that DAG would look like going through all the agents? and we’ll come back to this. Okay, back to our data set here. This is cool, right? So it knows I asked a question around what can we analyze? And now it’s going to. Tell me, like, what can I ask about it?

So let’s just see… I have a question I’m going to ask it already, but this will be interesting to see what it says. And so you can see it filled in that question for me, actually, right? I didn’t even have to type it or say it. It recommended that, like, that’s probably the next thing I might want to know. And so I just had to tap through. So it’s giving me a lot of good questions here. So funnel and conversion question. What’s our overall conversion rate by device trends and anomalies? How is revenue trended month over month? customer satisfaction. What’s our NPS score? And how does it break down by segment? Um… I like the, let’s go NPS.

because then we can make it explain what Mps is, too, if anyone knows what that is. What’s our overall NPS score? And so now now for the first time, it’s going to start doing some analysis right? And this is a very simple question. It should be fairly fast. We’re not asking for some in-depth thing about like, what’s driving NPS, or what are recommendations you would have to. Drive NPS. So it can do those things. So our overall NPS is plus is 42.3. And it breaks down by promoters, passives, detractors. to say like, hey, I don’t think everyone here might know what MPS is. Can you.

Uh, you know, give me the TLDR on… What MPS is, why it’s important, and how you calculated it here, like what data sources you used. And so I should do this for us now. I like to do this a lot, just to double check. Even if I know what the metric is, I like to confirm that it is aligned with me on the metrics. And you can store those metrics in a knowledge base. So it’s like. Constantly looking at that. But, you know, this is why I say it needs you to validate. So I do checks as I can. So yes, the MPS measures how likely your customers are to recommend your product to others. It’s a survey.

And so basically in the survey, I’ll get you say on a scale of 0 to 10, how likely are you to recommend us? I know, drop in the chat. Has anyone answered? survey like this from some product. They’re in the middle of doing something. and uh… the asterisk 10. Attendee is screen share still on. Can other people see my screen? Yes, yes. Okay, people can see my screen, and I think people can also have also answered NPS surveys. No, you’re good. You’re good. No worries. If anything breaks, like, that’s why we have the chat. Um, so MPS is calculated by you move the the the passives, and you do percent promoters by percent distractors. And then I like checking the sources a lot.

So what they use here, I created the NPS responses table in the Novomart data set. It has, yeah, okay, so it tells me about that table. And so, uh, something else you can do here, like, you don’t have to take this word for it here. Like, when I am validating my own data, I actually have it connect to other sources with MCP or API, which lets it talk to other sources like. Notion or Google Docs, or something like that, and it’s like, Hey, every time you query something or do a calculation, go like.

paste what you actually ran in there, because I’m gonna go validate it myself in the database, running it manually, or look at that SQL that you ran and actually look at my past analysis to see if it ties out. Okay. It really wants us to dig in. This is like the cool thing that I think’s been in the past few weeks, like, and because we. Um, well, I guess I’ll get into this. So while this runs, I’ll tell you a bit about it. So want me to dig in to any particular cut of? NPS. And I’m not even saying anything right now. It’s recommending I ask it to break it down by user segment, so I’m going to go with that.

The reason it knows how to think analytically right now is because when we built this, we built it within the repo of our course. So we have a course called. AI analytics for builders. It’s a five-week course, where we teach a ton of analytical framing, uh, analytical thinking, um. different types of analysis, how to do storytelling, like, maybe drop a one in the chat if you’ve came to any of our previous lightning lessons. Like, we had lightning lessons around. designing metrics that don’t lie and framing analytical questions. And so all these lessons that we’ve done as well as like our entire curriculum for how to, like. Think analytically is was in that repo.

Like I’m talking about like dozens of hours of content around everything that high and Stravi and I have learned in the past 10 to 20 years of our career. And like it basically consumed that. And it was like, and now it knows, like, what it should look for. And then we we took all that and restructured it into agents and skills in a much more. kind of like workflow oriented way. But that’s why it knows how to dig into this stuff. So it’s not necessarily going to like happen out of the box. You got to teach it your analytical framework, and that can be totally different for your use case and and your company, right? So here’s the NPS broken down by user segment.

So plus members are significantly happier. Their NPS is plus 5, 4 passes are nearly identical samples evenly split. want me to dig into the detractors and the free terracing and the carments, or trend this over time. Yeah, let’s dig into. the detractor comes and also. I went to. understand why they’re detracting. So. Let me know if there’s other data sources we should look into to validate. Right? Mps is like this kind of qualitative metric. If you’ve if you’re a data scientist like you’ve probably got the question before like, Hey, can you, like. tie our MPS scores back to product usage or back to so-and-so, and it’s like, man, that’s such a lagging indicator.

And it’s like a biased small sample. But let’s see if we can try and get Claude to do that. While it does this, I’m going to go back to our other terminal and kind of show you how this works. And we have slides on this, too, that we’ll probably go through. But maybe we can skip it here. Okay, so you can see it’s a very long DAG of like how this, why it can think so well. So like. basically a user for this AI analyst repo we have which is all open source. You can check it out, or like, I don’t know. You don’t have to use it. You could just like use it for inspiration to build your thing from scratch. You can rip it apart, whatever you want to do.

Ideally, go add to it and make it better and and push up whatever your your forked version is there to keep it open source and have more people benefit from it. So, a user has a question. It kind of does it starts doing stuff in parallel. So in parallel, it’ll start framing that question better. It’ll also start exploring the data and tying out sources to make sure the data is valid. down this slide from the question framing, it’ll start taking your questions and turn them into hypotheses. Then it’ll hit a checkpoint where it’ll verify, like, the framing is correct, so this will come back to you if you want, or you can tell it to keep going on through.

But this is where, like… Yeah, like, yeah, this is so funny. Skippable if users said just do it, because a lot of the times I’m just like, just go through. Just, you don’t need to ask me about the question. I know you got it. But it’ll stop here if you wanted to and be like, hey, here’s some questions I thought of. Here’s how the data looks. that I think we could answer these questions. Do you want? Which one of these do you want to do? Then it’ll go into descriptive analytics. So segments, funnels, etc. After it does all the analysis around descriptive analytics, it’ll go into like. root cause investigator. I bet you that’s what it’s trying to get to with this NPS thing, right?

I asked for our overall metric, and then it asked me if I want to segment it, and then it asked me if I want to drill into that one segment that’s going down. So starting to, like, figure out root causes. Then it’s going to go into validation. So like, do some structural checks. Business rules, Simpson’s Paradox. It gets business rules by, like, connecting all your other sources to it, like Notion, Google, whatever, Slack. Then it’ll start sizing out opportunities from what it learns. So it’s like, oh, this, like, root cause, this thing’s bringing it down. Um, if we were to fix that. like, um… What could that do for our business?

And let me tell you, no matter what data set you use, like, there is a story behind it, and there’s something going on. So, like. that’s the other reason I love this thing, is because, like, there’s just stuff where I think it’s business as usual, and I put it through here, and I might not show the deck or the output to my stakeholder, because I have to, like, maybe it goes too deep in one direction, that’s… Not tenable for the company, but it will give me, like, really good insight and understanding of what’s going on in the business, so I can ask it better questions for analysis later on. And I can think about it.

That’s what I’m talking about, getting less mechanics and more like thinking time. It hits another checkpoint, starts doing analysis verification, and starts like figuring out the story behind it. So this is a really critical part, and I think this is one of those things where separates like the good data scientist from the bad, not bad. Great. Sorry, good to great data scientists. We do a podcast, and at the end of every podcast episode, we ask, like, what’s the most underrated skill in the data profession? And it’s always, like, communication, storytelling. And so we have multiple agents that are on storytelling and narrative coherence and like narrative arc.

So it designs a bunch of narrative beats. goes through another checkpoint to, like, create storyboards. And then it starts making charts and visual design. eventually, you know, it creates a deck for you, right? Like we’re going to go over this in some of the slides. But you can kind of get the idea here. Let’s go back to our NPS analysis. Okay, interesting finding here. So. Detractor comments. The comments are… Yes, this is synthetic data. No shit. The comments are templated. It even knows the data is not real right? So yeah, had trouble with with payments. App keeps crashing. Check out confusing on mobile. And here’s I remember if you remember, I asked it to like look at other data sets.

So it cross validated on our support tickets from the support tickets table, which also sees delivery issues, payment issues, product quality issues. Delivery and payment dominate support tickets, too, which correlates with NPS. Um… One caveat, this is interesting. About half the detractor comments are actually positive. Love the product selection, great service. That’s so funny. This is the quirk of me creating synthetic data. The comments are randomly assigned and don’t always match the score. So it’s like, you can’t even trick this thing right? Like, but I’ll say, like, we’ve I’ve used this. We’re applying this at our our day jobs, where it’s not synthetic data.

And, like, it fricking works. So it digs deep deeper. Can you look over time to tell me if there’s like points in time where these issues are arising? And so I’ve ran through like different versions of like looking at some of this stuff, and. It’s just like any analyst, right? It’s not necessarily going to be like, I look at segment first, then I look at trend, then I look at something else. It’s going to do a bunch of stuff and it’s going to, like, try and identify, like. things going on and go down those different routes. So it’s not going to be like the same every time. So sometimes you need to, like, tell it where to look at a little bit more.

Attendee, I would recommend, um… Getting started in Claude before, and cloning that. I mean, you don’t have to, but like maybe clone that repo and, like, play around with it a little bit. The boot camp is going to be more about, like, how do you build your own thing? But we’ll definitely do some intro stuff on getting stuff in Cloud Code. We’ll probably spend the first couple hours on that. And we’ll also have like some pre-work to that people can read through. And, like, a recording on that. That you can do ahead, but I would definitely, like, get set up before. All right, what’s it doing? So it’s looking at detractor NPS comments by month. It’s looking to support tickets by month.

Here’s what the data shows. There’s a payment issue spike in late May, early June. The normalized support ticket data revealed a clear payment issue spike. outside of that spike trends are relatively flat. Okay? So it’s. It’s kind of getting into, like… stuff that’s hard for me to just read here, and, like, I want to pause now and be like, all right, what, uh… How do we visualize this, right? Like, so almost like. hey, I’m having… a difficult time digesting all of the learnings. Is there a. way you could visualize what’s going on over time. with this, uh, whatever this anomaly is in a chart that I can share with my team.

So I’m doing this like where I’m kind of hand holding it, going from step to step right where it’s not doing the full pipeline on itself, because I I just want to show you guys, like. If I went through the whole pipeline right now, it’d take a long time to get through every spot, and you get all the charts, and you get all the insights and all this stuff as output that you can look at along the way if you wanted to. But, um, it’ll keep running all the way through until it gets to a deck. And I just want you guys to see what it’s doing along the way, if we were just to talk to it. So let me create the chart. So yeah, so I just like. expanded. It’s connecting. It’s creating some charts.

I just saved it. Let’s see. It saved it in outputs payment spike timeline Png. Let’s see how. good a job it did payments back to him. Well, that’s pretty nice, actually. So here we go. So the reason it does like nice charts like this. is, um… I spend a lot of time like kind of nurturing it and growing it to, like. Have you guys read that? Has anyone read that book, Storytelling with data? Maybe drop a one in the chat. If you’ve heard of that book or you read it? I freaking love that book. Like, I that’s like kind of like my data viz Bible. sends, I don’t know, I think 2018 is when I first read it. And so the principles in there are like great.

And actually, I’ll show you how we work those principles in here. But like you can see, you know, this isn’t some, like. chart that just has a bunch of clutter on it and says like. I don’t know. tickets over time, right? It tells you what’s going on in the title. So it’s like you like, I don’t know. We could deliver this to like someone pretty senior, and they’ll know what’s going on like it’s very clear. It’s like, okay, there’s a payment issue that spiked. Uh, and Forex in May like it only looks at like it takes out the payment issue from everything else. Let’s try to make another chart. This is great. I want to have a chart to accompany this that. shows what we’re seeing in NPS.

I don’t even. So this is pretty vague, right? I just thought… I was gonna tell it more details around, like, I want to tie it to support tickets or something like, let’s just see what it does, though. Like, you can ask pretty vague questions because before it makes this chart, it’s going to go through the framing questions workflow again. And kind of figure out what I’m really trying to ask. Okay, I just made another chart. So, and then typically I’ll make the charts, and then it has another agent that like reviews the charts before, and it might like iterate on them a few times. and you see it’s also right. It’s also giving us information here that we can talk about.

So free users sit at 30 NPS. Let’s open up the chart. Free users sit at 30 MPS throughout the year, while Plus members hold steady at 54. Together, these 2 charts tell a story. The payment incident was a real. So this has free users. NPS score plus members. Is there any way to… Show how Mps and the support tickets are correlated, or. I want to basically show my team that. They’re reflecting the same issue. And we’ll just see what it comes up with. Yeah, you can make… you can have it make interactive web pages and dashboards. Like, I haven’t, like, right? I haven’t set the agent up to do this, but we’ve talked to, like, friends who have have done that. Like, the whole like.

The thing that makes this work really well is it has a ton of guardrails on it through the agents and the skills. So I’m not trying to be like cloud code. Build me a website on NPS. I’m trying to be like cloud code, follow your workflows and your agents that you have established where you know you can like. I’m confident that you can make this chart reliably and think analytically. You could add an agent that has a bunch of rules and instructions on, uh… how to like. Go, uh… make a website about this, but I I and you can try and just like zero shot. It usually like you want to 0 shot. See what it does, correct it, and then tell it to make an agent about it.

Usually I start with like a planning thing where I have, like. a bunch of, like, architect agents, like, blueprint out, like, what would be the best way to, like. Make a, yeah, HTML like page or something based on this that’s interactive. Yeah, you can ask, so what? Then what? Let’s try it. After it makes this chart. Uh, did it make a chart for us? Tickets versus NPS correlation. So, yeah, support tickets, might coincide with NPS dips for free users. Let’s talk… let’s try to so what? Then what thing? Okay, you’ve taken. You’ve you’ve… dove into this a bit like, what’s this? So what? Like, what should we do next?

And then what we do next doesn’t necessarily have to be what the business does next. It could be, what do we need to analyze next to understand more? So we’re doing like one analysis right now. We’re kind of going back and forth. Like I said, you could kick this off end-to-end pipeline. Where it does this all and follows this, uh… Workflow, uh, down here on its own. Without you having to… interrupted a bunch. And what’s cool with that is, like… Then you can just go kick off, like, 3 or 4 different analyses at the same time. You don’t have to be, like, just on one. And it’s holding all the context, so you get that out of your mind.

My recommendation to start with what caused the payment incident is the most concrete thread. Okay. We might as well do that. and I want to see what it’s going to think. So. It’s thinking. After this, I’m going to try and have it make a deck for us because you guys probably kind of get the the gist here. Let’s go into control O to see what it’s actually doing. So it’s checking a bunch of stuff right? So there’s a there’s a whole data set that experiments happen this time. So it’s checking if like. If there is an issue with experiment assignment, it’s looking at the payment tickets by experiment. I really want to see if experiments run. Hmm, I found something.

The app version is a smoking gun version 2.3 appeared during the incident window. with 465 tickets. It didn’t exist before that incident. So now it’s going to dig deeper into that. This is, like, the thing that’s pretty cool about it. It can go down, especially because like it can. pull up a bunch of sub-agents simultaneously to go check a bunch of hypotheses at the same time. So it can create hypotheses. Everyone, these sub-agents, go check each hypotheses, and then some of them are, like. Okay, there’s nothing here. Experiments looks good. We don’t need to dig deeper. And then it can look and see, like, Oh, something weird is happening here. like a unique anomaly.

Like, let’s dig deeper there. So it’s, like, really good at routing its thinking and like, yeah, I just like we can go down that stuff. Yes, many rabbit holes. And then I was like, how far. to go down those rabbit holes. And yeah, of course, we need to like tell it sometimes. But like, I can’t go down every rabbit hole. I have definitely been at jobs where I’m just like for months trying to figure out why our sessions are dropping and just like it’s stuck down a rabbit hole for like 3 days, and it’s just like, there’s nothing there. Um, but this can figure it out a lot faster than us, and it can do them all at the same time. Um, alright, so what happened?

Ios app version was deployed on June 1st. It immediately generated a massive spike in payment-related support tickets. 100% of the 2.3 payment tickets came from iOS. Okay, this looks like it’s the issue. Oh, I was gonna ask it to create a summary chart, and it already had. It already had that as like the recommendation. go off. So let’s see what it creates for us this time. And then after this, we’ll see if we can make it make a deck. The deck thing, I think, is the thing that takes the longest, actually, because it has to spend a lot of time figuring out the storytelling and like the narrative arc, and it has to like.

go back and look at, like, narrative after it creates the story that might recreate charts to tell the story. So it doesn’t just like look at the charts you have and then create a story off it, although it might do it this time because we kind of have been doing it step by step. And then I have to go back through and like look at narrative coherence through all your slides. So we’ll see. I’ll check this out. So this is like the cool thing, right? So I have our I have our reviewer. Um… I have a reviewer of the charts, so it says here that annotation is colliding with the subtitle. Let me clean it up. I wanted to get to it before it cleans it up. Nah, too late.

So, I bet you this annotation of 2.30 tickets was overlapping with the subtitle right here. And then the reviewer caught that and came in and fixed it. So, I don’t know, like, yeah, that’s, like, pretty obvious, like, the tickets are just coming from this iOS 2.3 version. The other Android and iOS versions, which are in gray back here, didn’t experience that ticket. And then it dropped right? Once, like, uh… There was a bug fix to it. Okay, yes, put these into a deck. I don’t even. Let’s see what it does. I don’t know if it’s just gonna like paste these things into a deck right now, or if it’s going to. Okay, it’s creating deck creator agent. This is kind of cool.

So you can see it’s kicking off the different agents. And basically, it knows that when I want to create a deck, the Claude MD file. I want to create a deck, and it knows to prompt. itself what to do. So one is it needs to go look at the deck creator agent. One, it needs to go to the presentation theme skill. We can go look at what those are. So. Let’s check out the… agents. creator agents. This is kind of what it looks like, right? So agent decorator, its purpose is to create a complete slide deck from the analysis, combining storytelling narrative with charts, applying a presentation theme and general speaking notes for every slide.

And so… It has like a bunch of inputs that it takes from other pieces of information we’ve had. So it’s not so like. The goal with like the agents and skills, I think, is to make it as, like, as deterministic as possible. So not just saying, like, make me a deck right now and having it try to figure out, but turning everything into like deterministic workflows, and just leaving, like, the reasoning to itself. So it’s like, you know, consume. consume stuff we’ve had around the format and the context and audience and and come up with this stuff. So non-negotiable defaults. So I have. I have multiple themes of slides that I create. There’s like a dark mode and a light mode.

So actually, you’ll see when we switch back, like, I’m using dark mode for this workshop talk. Claude helped me a lot. He probably did like 80% of creating the deck for this workshop. So it’s not just analytics. Yeah, tidal collision prevention recommendation ordering. Yeah. What are required HTML components? Valid slide classes. So it has lots of different types of slides. It could make that we’ve worked through so far. So again, like, Kpi slide, have a full chart side, put the chart in our right, put the chart on the left. I think that these companies like Gamma that are like AI presentation things are just going to get eaten alive by Claude.

seeing what it could do on 4.6, and… be honest, it’s even wilder. What did it do? So then it was reading Deck Creator, and then I was reading presentation theme skill. So let me show you guys skills. So… Agents are just, like, who’s doing the work, right? And then whoever’s doing the work just like us has skills. what’s it called? Presentation thing? There it is. So it generates projects that look professional, telecoherent analytical story, and follow consistent theme standards. And so this is like where the real details are. So in that agent. workflow, and it’ll tell it to look here. So like a title slide, like, here’s like a format of it, like.

A headline, a subtitle author, executive summary should have like a few findings. Context should say the question, the data method. Insight slide should have like a headline synthesis slide tells the so what, which we were talking about earlier, right? Um, and you can expand this as much as you want. And you work with Claude to expand it. Why so many special characters? So. Uh… Basically… basically combined in a lot of this. I mean. That’s just for readability. A lot of special stuff. So if I do the preview like this, this looks like a lot nicer. It was just like a markdown file. But now we can read it a little easier. Um, how do I create such extensive MD files?

I worked with Claude to do it, and I don’t just ask Claude, I, like, go into what a skill I call architect mode. And basically, architect mode is like, say, I say like, hey. This slide you’re making sucks. Like, it is missing the insight. Things are colliding. it’s like the colors are off. I don’t know like also, it’s not really providing value. Here’s some things I envision where I could slide to be like expand on my thoughts, and then enter architect skill mode, which in that case, it basically takes it spins up, it decides what personas could help me plan how to solve that problem. And so it could be. 2 personas or 7 personas. So it’d be like, alright, we need a storyteller.

We need like a. engineer. We need, like, a graphic designer. I don’t know. It’ll come up with them, and then all of those sub-agents will basically go create their own plans separately, not talking to each other. So maybe there’s, like, four of them. They all create separate Markdown files with their. With their solutions for how they would create this, like, agent. Or a bunch of agents, or a bunch of skills. They all come together, and they debate and argue around how to do it. And then they go back to their own markdown files again, and they edit all of their plans based on the feedback of the other sub agents, and they come back again and they argue and debate.

But the most important thing is they align. On a single master plan to create it. And then I took a look at their master plan, and. Sometimes I review it. I have to. steer them in a different direction. A lot of time is pretty good, and then I just execute it, and they’ll create the agents for me. So everything can build itself. One of the crazy things we did actually was we built this whole AI analyst in a few days, and then we. said, hey, if we wanted you to build yourself again, just like if we were to give you, like. Cloud MD file to build yourself. Like, what would that look like?

And it compressed like 40,000 lines of code into 1500 lines of code, into a single markdown file, which is like a seed or a genome of the entire AI analyst. We push that to Git, and then we pulled down on a completely separate machine that never seen anything like this, and we had it rebuild itself from scratch. And so it basically, like, we said, go look at the Cloud MD file. It woke up, like, four personas to plan out, like. Okay, let’s make the… let’s make the blueprints for how we’re gonna build the how we’re gonna build the AI analyst. Basically, like, they’re designing all of the sub-agents and agents and skills required to build the AI analyst. Then it created all those from that plan.

And then all those builders, like, planned how to make the actual AI analyst and went through and made all these sub-Asian stuff. So it’s like extremely portable. You can use Claude to build a lot of the stuff on your own, which is what we’re going to go in our doing our like boot camp stuff a lot. We’ll teach you how to do that, which is, like. like like I said, like, play with our version of it, but, like, you can build this, like, very specific to the context of your company to make it like. work far better, I think, than taking, like, generalized products and are buying some freaking expensive product that’s gonna do this for you. You can just mate yourself and pay for Claude code.

Um, which is probably cheaper. All right, uh, where were we? Yeah, what’s it doing over here with this deck? Okay, it says I made it. The deck is ready. So what we have at Output is called a MARPMD file. This can be converted to a HTML or PDF. If you do the HTML version, it’ll have speaker notes in it, or you can just get a Pdf. So here’s the slide structure. Okay, it actually looks… It looks like I went through a lot of stuff. I don’t know. We’ll see what it does. But at first, I was like, is it just gonna make a deck of our charts? Or is it actually going to do something? So it’s just output deck Pdf. Ah, it’s not going to work in here. All right. I got to go into another.

Gotta go into another window. Actually, let’s do this. I’m going to open it up in HTML. And then we can just go back to our presentation. Hey, can you actually make a HTML version from the Artmarp MD and then open it in browser? Right? So you do… this is this is a cool thing, too. You can just, like, have it open stuff around your whole environment like you don’t just have to have it living here. So, like. I will literally be like, hey, can you? I just got this question from. Uh, this person in Slack, because you can kind of just like, can you like answer the question for me like give draft it up like maybe they’re asking for like.

Yeah, what’s our NPS looking like over time draft a slack, and then can you just go send it to them? Okay, let me, um… switch my screens. We’re going to go back to the. presentation. But as part of that we’ll go through our slides. Okay, so here’s our slide deck it just made for us. That’s pretty quick, right? We were kind of talking about a bunch of other stuff. I don’t know when it actually finished. So Ios V. 2.3 broke payments for free users. How to do? Oh, okay, I think it actually went through the full side deck process. So free users are less satisfied. So the first thing it does is it sets you up with context. It’s really important, right?

Like, you can’t just, like, dive right into those graphs. You have to tell people what’s going on. So plus 42 overall NPS, plus 33 free tier. blah blah. The problems of structural NPS gap between free and plus users. Mpsgap free tier satisfaction lines plus member year round. So here’s our our slide from before free to MPS Trails plus members by 24 points all year. This isn’t a one-month problem, it’s a year-round experience. So, talks about what they’re citing. This is a lot of that information. It was returning us on Slack, right? Not in Slack, in the terminal output. It’s like, I feel like I’m talking to it in there.

Something happened in June payment incident, single Ios release caused the 4x spike. Here’s our our graph there we saw before. Both support tickets and NPS reflect the same friction points. One device, one version, one month. So, you know, tune it to. That’s kind of intense. But tune it to your voice. I don’t know. This feels like a beginning of some like trailer of like a Steven Seagal movie or something. one device. And then it gives like the the proof of that. And then it shows what the evidence points to. And then here, now it’s getting us into so what? So 3 actions to prevent this from happening again. So yeah, conduct a postmortem on ios v 2.3 with the mobile team. It’s a good idea.

Add automated payment flow monitoring with alerting. Uh, yeah, the Cincident ran for three weeks before I shipped with sick should have been caught earlier. Investigate structural free tier friction beyond the incident. So there’s already, it’s saying like there’s already a gap between those two. So there’s other stuff probably going on, too. Alright, let’s get back to this deck. Oh, and I’ll just show you this real fast. Like, why did the conversion rate. We just went down a totally different route. But like.

The question I had originally planned, and I… basically, I ran this question last night, and it took, like, 45 minutes to go through, so I didn’t want to like run the whole thing and kind of bore you guys, although we just spent 45 min at least on that. I asked it just like, why did conversion rate drop 4 points between Q1 and Q2? in which customer segments drove the decline. And then this case, I told it, run the whole pipeline end to end. Don’t. You don’t need to stop for me. And then I went to sleep. Like, I ran that at 8 Pm. And I… Glenn and I did my nighttime rituals. It wasn’t quite done by 8.30, and I was like, whatever, I’m getting in bed.

And in the morning, it had my deck ready for me. and you can see it falls like. Similar kind of things. gives us the context. So conversion rate dropped sharply from Q. 1 to q. 2. The aggregate funnel shows degradation at every step of the funnel. At first glance, it appears to be broken across the board. But if every segment got worse, how can the client be a statistical illusion? And it says, like, basically, most of the decline is caused by actually user based growth. So it’s like a mixed shift issue, not an actual funnel issue. So this is like Simpson’s paradox. This is the one thing I have it checked for all the time now. Simpson’s paradox comes up like.

all the time in analyses where, like the overall trend is the complete opposite of the segmented trend, and that is just a sign that something is going on. You should drill into, and it’s probably not going to be what you think it is. Anyways, you guys kind of already saw the last one what it could do. So I’m not going to go through the whole thing, but. Yeah, in this case, it actually does the opportunity sizing around the cost. And it starts giving, it gets a bit more into the opportunities. and it explains the math between behind that. And yeah, gives you some actions at the end. All right. Let’s go back to here. We’ll… go through some more slides. And then.

I think Sravya and Haya have been answering a lot of questions as we go. I have not been looking at that. But, uh… we’ll definitely have some time for questions, more questions at the end. Alright, so… I’ll skip over some of this, because we kind of already said it right? So it goes through question framing. Then it goes into. Data exploration, so it figures out, like, it doesn’t just, like, go look at every table. It like looks at the question and then it goes and sees what’s all the data is available. And then it picks like in this case, like it didn’t go look at 13 tables. It might pick like three or five.

because otherwise it’s going to take way too long, and there’s probably a lot of tuning that could be that could happen on that. They write SQL, it runs it. It’s in this analysis stage. It identifies quality issues and real patterns. It decides what’s real. Uh, does the validation, that’s where it picked up, like, Simpson’s Paradox. It is, like, structural validation. Like, do the numbers add up? Are they logical? Do they they segment to total? It will look for, like. crazy anomalies that. make… don’t make sense like if every single segment plummets like 300%, it’ll like flag that as like, I bet there’s a data tracking issue.

Um… It then creates the story and output, and, like, so if it hits a bunch of stuff in validation, if it finds, like, if it finds, like, hey, Simpson’s paradox or structural issues or something, it starts initiating warnings and blockers, and it’ll go back and try and, like, dig into what’s going there. It won’t just, like, go through and give you the analysis. It’ll try and, like, loop through and find as many like. things going wrong. and redefine its analysis based on that along the way. Yeah, so rights narrative structure.

It starts with an executive summary first, st as you saw, then it gets into the evidence, and then it gets into, like, say you mentioned breakdowns, and then recommendations. So it’s designed. It designs the story arc, and then. And in our case, we actually built a bunch of charts by hand first by hand. We handheld it to buy build charts first. So that’s why it used those in other cases it may build charts, and then. not build charts at all, it’ll probably figure out the story arc and what’s going on, and then it’ll build the charts up best do the storytelling. after the fact. So, and what comes out, like… Yeah, to some degree it depends how simple your question is.

Sometimes you’ll just get a stakeholder ready analysis out, and you could send it. Other times you’re going to get an analysis that you should look at. In terms of forming more ideas of what analysis should… what the analysis should be. No matter what, you need to validate the numbers, like, before you send anything out. But, like, sometimes they’ll tell a story, and it’s like, oh, that’s, like, not really what my audience is looking for, that’s not really what my team needs. And… you know, you can pick and choose, you can go manual mode and make your own thing then, or you can, like, have it rerun with a with a different goal. is a perfect like no.

you’re going to want to adjust the framing of your… for your specific audience. Like I said, you might want to reorder things. You’ll definitely want to double check the recommendations against the context that only you have. Um, I almost think, like, the recommendations slide as, like, good ideas for you to use, but I don’t even… I… so far, my recommendations, I haven’t, like, given them all to my stakeholders. I’m kind of like, oh, that’s a… Crazy recommendation. You want to spin up a new dev team to build something? Like, I’m not going to recommend that. But I mean, I hope you’re starting to see this that like. It does a lot of like. Like, that’s that 20%, right? Like.

of thinking that you do… like, that’s a 20% of mechanics work that you still have to do. But, like, the 80% around, like, SQL and, like, some sorts of simple validation, chart generation, narrative structure, like, that could be done. Um… Things that can take… Days could be done in, you know. payday or hours, or minutes. And like. I don’t know the the the validation part’s what I really like, that it will, like, stop itself and go and validate and rerun things automatically like before I even see the output, and then they’ll tell me that it did like that’s that’s. That’s the difference between like a correct finding and a misleading one that I think was not happening in previous models.

Uh, let’s talk a little… I kind of actually walked you through a lot of this stuff, but we can say some things about the fact, how cloud code works real quick. So applaud Code MD. It’s like, this is the single most transferable pattern from everything we’ve built. Whether you like use our system or fork it or build something completely different. It’s where you want to start. Claude MD. It’s file that sits in your project. reset the start of every session tells Claude, you know, who it is, what to do, and most… and you can tell it to learn from previous corrections. That’s like the very powerful part. Um, you read it once and then you gotta keep refining as you go.

And every correction gets added and you can, you can set up Claude. that to have a skill which what we’ve done here is like, Hey, when we tell you you did something wrong, you corrected, like, add that to the MD file. And the next section I’ll know in advance. Um… It’s a quick architecture overview. I’m going to keep this pretty high level. Cloud MD is your map right? It points to agents. Agents are the workers. One frames the question. One runs the analysis. One validates the output. One writes the story. There’s actually could be multiple agents, sub-agents under that that then tell it how to do each of those things. Like they’re not fancy. They’re prompt templates with rules.

Each agent has skills, skills are. behavioral guardrails that auto-trigger, so when analysis agent writes SQL, a skill checks whether it’s using the right date columns. When the validation agent runs, skill enforces, like, that four-layer check. Uh, you don’t… you can invoke skills manually, but, like. In these cases, they’re not invoked manually, like, they fire when they’re relevant. And a lot of that’s baked into, like, the… the MD file or the agent. Everything that touches a number goes through Python or SQL. So the name of the game, I think, is make non-deterministic systems as deterministic as possible. So… The LLM can reason through the analysis. It never does arithmetic.

That’s how you get deterministic output from a non-deterministic system. That’s how you, like, reduce or remove hallucination. Um, and then there’s knowledge. That is what compounds. When you correct a mistake, it gets logged. When you define a business rule, it’s stored. When you connect tools like Notion or Slack or your Data Warehouse, or Google Drive, uh, Google Workspace, like, that context can get folded in. In the next session, it knows more than it did last session. And it is extremely good at connecting business context to like query history and data context. In my experience thus far.

So… This matters because most tools give you like the same workflow for every question like if you like, I don’t know. I feel like a lot of like solutions out there are just gonna like be like, we’re gonna look at this, and then we’re gonna like this like a quick data pull doesn’t necessarily need a 19-step pipeline. So here, like the system. and we can make this even better, right? But like we’re trying to make it so it classifies our question automatically like level one’s a look up how many users last month, like when we asked about the Mps. one query, one number, and it’s done. Level 2 adds a visualization. Level 3 is a light analysis, like segmenting something or charting it.

Level 4 is a full pipeline, level five is a deep dive that might take an hour or more. So you don’t necessarily need to tell it which level. You can just talk to it like you would an analyst or a regular human, and it figures out from your questions, like. A quick question gets a quick answer. A hard question gets a thorough investigation. The routing isn’t just speed. It’s about appropriateness like not every question deserves a deck, right? Uh, Sravya, what’s up? Hey, Shane, I was wondering we are closing to 11, right? I was wondering if people were going to leave, and you know, how are we doing on the time? Just a quick time check, yeah. I think we’re good till 1130 is when this goes over.

Yeah. Oh, okay. Okay. Thank you. Okay. So I want to be clear about something. You don’t need our system. this up like we’re not selling a fricking product here like this repo is open source. You’re welcome to it. The value isn’t in the code. I honestly think, like. a lot of products. that are there today in the world. I’m not just talking about analytics, but everywhere like are going to get like terribly disruptive. So, like, I think it’s like gonna be turning more way more into, like, what can you build yourself and set up that’s specific to your use case, rather than like. purchasing solutions. I don’t know. We’ll see what happens.

But, um… like, it’s more in these 4 patterns, I think, that are key. Like, write the Cloud MD. Veterans Claude code into an analyst for your domain, like tell what data you have, how you define your metrics, what mistakes to watch for. This alone will totally change your workflow. Number two, number 2. Add skills that enforce standards. So when I write SQL, make it check the date columns. When it calculates a rate, make it verify the denominator isn’t 0, like, you know. Those are just python… those are skills that point at Python helpers. These fire automatically. You can just set them up once. Like simple rules, like any of us know how to do.

three log corrections, so when it picks the wrong column, tell it like you just freaking tell it like you would tell a person. The correction can get stored. You can talk to it like how do you store this? How should we set this up to store it next time, whether it’s a YAML file or a JSON? or something we set up in our data warehouse, or like, do you think we should build a semantic view of this instead of, like, having a bunch of tables? You can talk to it about that. It can help you. Next time it checks the correction log before writing the query and it has a much lower probability of making that mistake again. Um, so you correct that wrong column once, and hopefully it doesn’t happen again.

And then for like make validation a step in the pipeline, not something you do after. So as much of the stuff validation that you can automate like tying out from different tables. Having a check of things, like, make sense, like, having to go check metric definitions, like, before you see any output, have as many layers of checks, like we have four here that already run, and if they fail, the pipeline halts and it’ll, like, reassess. the analysis, or tell you what’s going on. So you don’t get like. You don’t want it to run the whole thing and get like a wrong answer with a caveat. You want to get a flag that says, I found a problem. Here’s what the problem is.

Like we should investigate before going with the same analysis. Then you’ll get like weird wrong results that you’re going to present to your stakeholders. The differentiator from just asking Claude a question in browser, right? Like. or ChatGPT or whatever, like chat forgets. This doesn’t. You run an analysis. You review the output. You spot something wrong. Maybe use the wrong date column. Maybe it defined active user differently than your team does. You correct it. That correction gets stored in a file. could be a markdown file, could be a YAML file, could be a JSON, could be something else. Next time, before it writes any SQL touching that table, it reads the correction log.

Because you have your Claude MD file that basically maps it the direction to look at that first. So it doesn’t make that mistake again, unlike chat, which is trying to hold everything in context and we’ll just forget. This is another critical thing. The most… the more you can optimize your workflows through your agents, your Cloud MD files, your agents. the less context you have to hold, the less it has to read your whole conversation over and over again, the less money you have to spend on tokens you can like minimize your token account this way. But I would say for the beginning, like, don’t even worry about that. This stuff’s going to, like, make you so much faster that like.

whatever the cloud code pro subscription costs like is worth it. After a few rounds, the correction rate drops fast by the third or fourth analysis on the same data set. Basically, I feel like knows kind of like that that small domain pretty well, and that knowledge just compounds. And then it’s nice because you can. work with it, to be like, hey, we’re gonna go into another domain now. Let’s think about all of the gotchas and things we had to learn to get our last domain. up to this level. Let’s not make those mistakes again. So so if you if you treat this like software, this is like kind of one of the crazy mind shifts. If you choose like software, you’re going to be disappointed.

So software should work out of the box. You should be able to install it, configure it, and it does the thing. So this is more like onboarding our really smart. junior analyst, I guess like 1st day you’re you’re correcting a lot. First week, a little last by week two, it’s producing work you might, like, put your name on and you know by week three or week four, it’s it’s catching things you missed. It actually like week one, it caught stuff I missed. I was like, why is this number off? and I had, like, incorrect abbreviations and like one of my queries for like joining tables. It was like extremely complex like 10 cte 15 and join table very annoying thing.

Uh, that shouldn’t even be query, it should just be like. some, like, aggregate table I had put together. But just haven’t had the time to, and it caught a bunch of my errors. which is pretty cool. The difference between the people who like, I think, get massive value from this and the people who like bounce off it is the people who get value treated like an analyst. They talk to it. They correct it. They tell it what it got wrong and why they don’t. They don’t file a bug. and to someone else to fix it. They have a conversation and fix it theirself. The quality of the output is a function of the quality of the conversation.

If you give it like super vague questions, it can do somewhat well, but priming it as much as possible, like. We’ll get you better results. So vague questions. You could get vague answers. If you tell it exactly what you mean by conversion, exactly which big column to use, and exactly how your team thinks about segmentation, especially early on. Then it gets specific very fast, and there’s less correcting you have to do down the line. Um, this is the thing… I want you to walk away with, which I think kind of clicked for me in terms of, like, how you use something like this. Um, to go from like. to basically not get your job replaced. So, like. I do it.

I also do an eval course, and at my day job I I I create a lot of AI evals for for our product and legal tech where I’m evaluating output that lawyers get. And it’s really there’s a lot of, like. challenges in doing that, because in order to understand. The output, like, when I see a contract. Or I see like some document that is a suggestion that a lawyer should do. To me, I’m just like. Looks like legalese like looks good to me. I’m a data scientist. I’m not a lawyer. But to the lawyer, they’re like, Dude, this is like really good, or like, this is not where it should be. So like if you’re a data professional, like you know your domain, you can do the same thing.

Like when I do it with lawyers, I have to like build a bunch of systems, and I have to like figure out how I can get all that feedback from them. When you’re doing it yourself. For a domain that you have experience in like in data you’ve worked with for like quarters or years that you know by the like the back of your hand, like you can move extremely fast. Like the only person who can automate you basically is you. And that reason is really simple. It’s like you’re the only one who can validate the output at the speed and quality. that you can, unless there’s someone else doing the exact same job as you.

So when the system picks like a wrong date column out of 5, and it will, especially early on, you catch it in seconds, because you’ve been seeing staring at that data for months. Someone else comes into your domain and tries to do that. They can’t. If I tried to go. Automate Stravia’s job? Like, I could not do it. I would have to pay Stravia to sit next to me and tell me if the answers are correct or not. I would have to, like, work with her schedule and build systems for her to, like, get that feedback in there. Like, I… but I can do every… I can validate all of my stuff, like. Super quick. You can catch in seconds. you like, I just can’t do that for you. You can’t do that for me.

So that’s why each person can, like, automate out themselves. So it’s not about replacing you. It’s about making you faster. You’re still the one who knows the answers, right? You’re still the validator. The system handles the mechanics, so you can spend your time on the judgment. So… Like, the progression, I think, looks like this like day one. you get immediate value. It’s like, or maybe it’s a little slower, like, it’s faster at writing SQL by hand, even with the corrections. Week one, those correction compound. It’s making fewer mistakes because it learned from yours. Week two, like, knows your domain. The output starts looking like something you’d ship.

And then, like, the real shift happens. You stop executing analysis and start solving problems. And so this is the cool thing with this, like, you don’t have to, like, go take two weeks. and automate all your stuff out and build this, and then you’re not gonna get 2 weeks of work done. Like, there’ll be a very, very small period of onboarding. of learning how this stuff works. And the only way to validate it, the only way to grow it correctly is to. make it do the analysis you were gonna do anyways. And so you instantly get value from it, and then you instantly speed everything up, and you start getting all this free time in your weeks, and then. like, you’re delivering all this value.

The rest of your company does has not automated their stuff out yet. And like only you can validate it. So other people can’t come in and use the analyst. So you’re just like. What do I do now? And it’s like you could go a different. You do a bunch of different things. You could make it more sophisticated. You could, like, go start solving. It’s going to come up with recommendations. You could go solve them, like, end to end. Like. I think, like, companies will get this wrong and be like, okay, go do it for someone else’s domain now. You go do that, and, like, that’s when you’re replacing people and automating yourself out of the job. I think you get extreme agency, and then.

You take that time, and you start to think and strategize how you want to use your free time. Do you want to do analysis on other things yourself? Like. What do you want to take your role to after you’ve, like, automated a bunch of these mechanics out, and you’ve flipped from 80% mechanics to 20% thinking to 20% mechanics, 80% thinking? So the repo is open source. You can clone it like after this or right now, like literally. Uh, I would say, like, start with… a CSV or like something simple like a query. You already know pretty well and feed it that like something where you already know the answers. ask it something simple, like. Like, what’s the average order value by a month?

Check the answer. If it’s right, great. Ask it something harder. If it’s wrong, correct it. tell it what went wrong and why that correction gets stored, and then ask the same question again, and see if it gets it right. So that’s the loop. Start simple, create what’s wrong, ask harder questions. System gets smarter each time. Yeah, so… We’ve been kind of working on this for for months. It didn’t really really work this. to the level it does now until a few weeks ago. The patterns, the architecture, which corrections matter, how to structure validation. We’re gonna have, like, a compressed boot camp in two days next month over a weekend.

Uh, we’ll, like… We can run through like some of the stuff with the synthetic data. But I would say, show up with your own data. It doesn’t have to be like work data sets if you’re not comfortable with that. Like, we’re not going to ask you, like, share your screens or your work data. But if you want to do that privately on on your laptop, that’s cool. I mean, I think if the stuff. It was pretty fun to do with, um… public data sets. So like, bring your own data, your own tables, your own metric definitions, and like.

Saturday, Sunday, I think by Sunday evening, you’re gonna have like the beginnings of a pretty working system like you should have your Cloud Md written for your domain agents that know your schema. or at least like the schema within like part of your product that you care about skills that enforce your definitions. And, like, we’ll start getting into validation that captures, like, specific to your data. So those four patterns, basically, from one of those slides I had earlier on. So I think if so, yeah, Surabi or hi, if you want to. You probably already did this, but drop the link in the chat. Um, for that, you can check out AI analystlab.ai.

everyone here, or if you’re watching the recording, I guess, like we have a 20% discount code. Claude 20 for the boot camp, something else for the boot camp. I didn’t put this on the slide. If you do the boot camp. So the boot camp is just going to be about like building AI analyst and cloud code. We have a 5-week course that goes through, like, the whole… the whole end to end workflow of like how to like leverage AI for making data-driven decisions like this. This isn’t actually just for that one’s not just for data scientists. It’s for. It’s for anyone to learn how they can leverage AI. We’ll have Claude code in there as well, but we’re not going to go into like the depth of, like.

the boot camp. We’ll probably share the repo, and people can leverage that in there. But we won’t be going into how to build your own. Uh, that’s all to say, if you do the boot camp, and you want to do the 5-week course later on, we’ll just subtract. Whatever you paid for the boot camp off of that five-week course. So you can just, um… kind of like pair them. You can do both for the cost of the five-week course, but you can just do the boot camp first to see if you… if you even want to do the… the full thing. Because some people already know how to think. Uh, through that whole analytical frameworks, and they don’t need all of that, right?

They just… they just want to take their skills they’ve already learned for analytical thinking and apply it in Claude Code. So that’s why we’re doing this bit separately. Um, yeah, scan the QR code, take out your phone, scan the QR code, that’ll get you to the boot camp. It’ll take you straight to the registration page. If you have any questions, like… We have a Slack community. Did I put that on here? No. It’s on our website, but I don’t know if you if Sravi or hi, you want to drop the link? It’s like bitly slash AI-connect. Talk to us in there, hit me up on LinkedIn. You can email me too. Um, yeah. Like I said, repose demoed today.

If you don’t want to pay for this, like, uh… Go get your company to pay for it. I gotta tell you manager Maven actually makes a template for like what you should email to them to like. to have them sign off on the budget. Or if you’re a manager, have your team come take the boot camp. Yeah, we run free workshops, too. So if you go to aanlslab.ai, we have some other free workshops coming up. Yeah, man, I had I had specific times. I was going to tell you about specific free workshops during this, and I forgot. We’re gonna do all right. So you saw how did some cool charts in there, some very nice charts. We’re going to do a lightning lesson. It won’t be as long as this one.

It’ll be like an hour. We do like 30 min. And like a demo and 30 minutes of talk. We’re just going to focus on data visualization in Claude Code because that was something that it did now that it really couldn’t do before. That was pretty cool. So, um… Yeah, maybe hire Shravia, if you want to drop the. link to the data visualization one. Yep. And, uh… Yes. Also, I had… I added another lightning lesson on building presentations and storytelling with Claude Court. Yeah, so we’re going to have two of them. One is the data wiz that Thai shared and the link above that is oh, I didn’t share it yet. I’ll share it right now. This is the. Yeah, thank you so much. I think I shared the right one. Yeah.

Nice, yeah. So, like, it can make pretty good decks. Like, you saw that one made. It made this deck, right? I know it’s pretty good. The first time I made this deck, it was like 45 slides, and I was like, cut it in half. We do not have that much time. But, uh… Yeah, we’ll have a whole free lightning lesson on kind of some of the functionality there, too. Okay. I think we answered some questions while we’re going. We have 23 min left. What other questions do we got? I think we have a bunch of questions. I think, uh, one of the questions that came up quite a bit is, um, where do we think this is all going? In terms of analytics, how people are… Mm-hmm.

what people should be studying, uh, in school, what should people be learning in the capability of what we’re seeing today, that kind of stuff. Yeah, so here is like… my philosophy on this like this is. your life. You got to take charge of this right now, because like there’s no playbook on this. There’s no. This is how it happens like. Don’t freaking let the CEO or whoever like tell you like where this is going to take you. So that’s why I really push like the validate yourself automate out, get thinking time and decide different directions. I think the 1st thing to do like. very short term is going to go like people are gonna build out agentic systems that automate out their work.

People are going to invest in the architecture behind that for like they’re gonna really clean up their data foundations. So that’s these things sing. And people are going to validate across companies. I bet you a lot of companies are going to actually hire more data scientists. I don’t know. I bet you they hire more data scientists to start doing a bunch of the validation or software engineers who like kind of get a little automated out of like building the product. become the validators. That’s some near-term stuff than I think other paths could be going becoming the solver. It’s like actually… Let me drop a link here. I read an article about this. last week.

It is a little… Yeah, I don’t know. You can check it out. I feel like it’s a little doomer, but it’s like optimistic. Of, like, where I think it could go. I think you could become solvers where you do the whole end to end workflow where it’s like you go through all the analysis. And then there’s going to be so many recommendations coming out from you if you take charge of this that you’re going to swamp everyone else with like really good ideas. You have to, you know, make those ideas digestible and actionable and like, not just throw a bunch of clutter at people. But there’s going to be some point where it’s like, okay, I did all the work that I used to be able to do, sent that out.

I did all the work that I always had to do, but never had time enough to do. Sent that out. So now you’re like five X. what you were doing before. No one can ask more of you than that. So now you get all this free time to be like. What’s all the problems and stuff that I never even thought of that like never would have even arised because they can only arise after I get done with this work that I could never have time to do. I think that’s some of the work that people are going to start doing. Maybe it’s more analysis. I think it’s more gonna be like solving problems and building things. I think there’s gonna be roles around like translators around the company.

Like if you get really good at this, like. just people that need to connect the dots, especially at big companies. I think there’s going to be a lot of people who are leading organizational change in their companies. But like at the end of the day, like, yeah, my whole take here is like just lean into these tools like this is happening. We’re on the frontier of something, and we got to own it and like dictate our own future rather than because if we don’t do it now. I think data scientists aren’t going anywhere soon. you know, like tech will probably, like, be first to replace some stuff, but, like. My wife works in the hotel industry and like.

you know, they barely use Chat Gpt like they got years. The business analysts there safe for a long time. Like, they go out of jobs for a while, but I do think, like, at some point, it’s going to be someone else. telling us, like, what those jobs are, or if the jobs don’t exist anymore. And, like, we need to define that by leaning in that. I think these next 6 months. If you lean into these tools could change like the trajectory of your career for the next 10 years, I bet. I don’t know exactly where that’s going to go. You can read that post, that’s like some of my early stuff. human resistant gorilla warfare strategist.

Yeah, I mean, and then someone’s gotta, like, make sure the, uh, data centers are cool. Someone’s gotta grease all the machines, someone’s gotta get the hamster wheel to make more energy. It’s like, humans are good for a while. you know, I… I don’t know. I live in Tahoe. We have a lot of wildfires here, like, the climate change isn’t getting any better, so, like, fires are gonna be hot in the future, so you can always be, like, a firefighter. Um, I’m joking about that. For the data profession, I think what I would learn right now, though, is. Learn how to build agentic systems.

that automate out workflows, especially workflows that you’re an expert in or can validate extremely clear quickly by yourself without having to have a bunch of other people validated in it. Get yourself time. And then the thing I would be learning is like evaluation stuff for like when you want to validate other people’s workflows, which I think is another path you could go. But mostly it’s just around like trying to figure out right now. I think how to build agentic systems. I don’t know, hi, Sravya, any additional thoughts on that question? Covert. I do want to go? I can. So. I think it’s going to be very interesting to see where we end up.

This is something that I’m literally trying to do with my team right now as well. How… we take all these AI systems and, you know, move into a place where what is the actual value add that we get in as a data professional. And. How can they, you know, find value from what type of analytical thinking we give? Like, what type of builders we become? Because we, like, I think Shane mentioned a bunch, he also mentioned in his article, please give it a must read. He shared it in the chat. about how all of our roles and, you know, the outlines are going to, like, just getting marched up and, you know, getting into one builder role, probably. I myself am looking forward.

The only way that I can contribute into this area is just being on top of my game when it comes to AI and everything related to where the industry is going and ensuring wherever I work, like my team right now, I have a team of 15 people. Ensuring, like, all of them, uh, also keep doing this. Yeah, that’s what we could do and see where everything unfolds. Okay, I’ll quickly share my thoughts on this. Um, I think the… Uh, especially for the new folks who aren’t, uh, super deep into the area yet.

I think there’s, uh, the… I see it as there is, uh, kind of like two things where, uh, the foundational fundamentals, so the ability to sort of, like, reason, the ability to understand logically, Um, like, things like that, is… almost equivalent to, like, the multiplication table. Like, we still learn it. It’s been many centuries, and that’s the foundation of everything else. Calculators come on, computers come on, doing the actual thing does not matter. much. But then the foundation of, like, you know, what is 2 times 2? 3 times 4? That is still extremely fundamental. So there’s the fundamental aspect of it, that… we need to clear, like, all of us need to, like, that… cannot be replaced.

Now, whether AI or, you know, like, or coding, or, you know, writing SQL or stuff like that is more akin to driving a stick shift versus multiplication table, like, no one really knows. Uh, at this point, build up your foundation as much as possible, don’t get married to the tools. get as ahead as possible with the latest systems and AI and stuff. So that you can marry these two, and you’ll be really powerful. That’s my thought. Thanks, Raviol. Alright, so there’s a really cool one. Uh, was there any concerns with stakeholders seeing the analysis produced by an agent? Okay, so… So here… so here’s the thing. That’s what you’re the one.

You can tell them it’s produced by agent, but you don’t have to like my whole ethos on this is like, you are accountable for your own work. This is the whole thing where I say, like, you validate yourself. This is not about. Stakeholders don’t get my agent. They’re not allowed to use it basically as my thing, because I don’t know what it’s going to fucking say to them. Like, I have to validate everything, but I can start validating stuff very quickly when I work more and more with it. So my thing is like. I’m going to keep doing my job. I’m gonna do it crazy fast, and I’m gonna go do other jobs.

So when I get every number, I have all those queries sent to a Notion page, where I can, like, check everything, I can run my own analysis. I can know… I can double check to see, like, the historical stuff to see if, like. That number is trending how as I expect it. And then I send them a deck or a change in the deck, or a number, like. I don’t know. It’s like it’s kind of like, hey, was your stakeholder mad that, like the analysis came from you you doing math on pen and paper versus like using a calculator versus using Google sheets. We’re just using a Python script. Like, that’s how I see it. This is just another tool that we use, but we’re responsible.

We’re accountable for every number it outputs, for every insight it outputs. When it gets it wrong, we’re not allowed to say, Oh, the AI got it wrong. Like, no, dude, like we’re using AI like just like we’re using a calculator like we can say, like, I. messed up something when I leverage the AI, just like, oh, I like. fat-fangered, like, thing on a calculator, like, we can say that kind of thing, but… when I give it to the stakeholder, like, that’s my mentality about it. and like so far, like, yeah, there’s not resistance to it when it’s framed and actually done and executed in that way. If things start, if we just give these tools to.

stakeholders, and this will come eventually, like, this stuff… In the years to come, or whatever, or sooner, I don’t know when. Like, something will be built where they can just go in and ask any question they want. It’ll take a lot of validation to get there. But, uh… like there I’ll see a lot of errors and resistance like I’m not trying to go down that path, basically. Does that make sense? Mm-hmm. Oh, yeah. The question was actually more specifically about, like, I hear you, and it’s not that I have a concern that a stakeholder is going to be like, ah, you used AI. Not good. It’s more so like they are used to seeing outputs from our BI tool.

So if I use the chart from this, then it’s clear that it wasn’t from the BI tool and it’s not even like they’re going to then be suspicious, but it does mean that they cannot go play with it themselves. Yeah. So I release this to my organization on Monday. So I haven’t had too much time to because it’s already new, right? So, but… I was honest with them. I told… I told them… well, I just told you guys right now, how I think about it, and then. I also told them like. Hey, by the way, like, you should probably do this for your role to like try and automate out your. your own workflows, and then because… and I told them, like, I can’t, like… I can’t.

I can’t like automate out you, or and that’s like, I wouldn’t be able to use like the product manager AI to manage this team like that’s something you would use to it. And I think that resonated pretty well. I will say, like, my company does not have a great BI set up right now. So it was for me, it was kind of like. Oh, this is this is great like our BI stuff’s kind of bad. If someone, if you have like a really robust system set up, I could see resistance there for sure, like, if someone’s just doing self-serve. No. I can add to that, Attendee. So Attendee, I work at a place where we have Databricks dashboards and stuff, right? So we could build systems.

So this is, you could think whatever we shared today as like a skeleton, right? This is skeleton of what’s possible. So, you could build MCP servers to whatever BI tool. I have a Databricks MCP server set up on my company right now. I’m working with it, and you could actually automate dashboards as well. So, which means that using this skeleton, this setup of agent skills and, you know, helper files, like all of this. We could do what you’re doing today and replace it with, like, a cloud-code automated system as well. So, uh, like Shane said, this is Opus 4.6 is two weeks ago, 3 weeks ago? I don’t actually remember. We just hit it on the week one when we saw this.

So, yes, there were earlier systems based, there were systems based on earlier models, but 4.6 is almost like a game changer in it. And I can see everyone’s adapting to this. Once we get to that place, every BI tool out there will have an MCP server attached to, you know, and once we get there, you could. literally do the same workflows with Claude Code. Okay, that makes sense, thank you. Great question. What other questions? We have 10 min left. Yeah, there’s a bunch of questions here. Uh, let’s see, so, uh, Attendee asked, one thing I struggle with doing analysis with Claude is when I want to twist a small thing in the result, but it had to fire the whole process again.

So, uh, how do we… Is there a work… is there a solve for that? I wonder if we can… what I do. I’m not going to switch screens. What I do is when Claude is going through Alex, my agents are set up since it’s all separate agents doing different parts of the workflow. I constantly have them write all of their. results for each step to markdown files include or Json files, including the queries they run. I have them solve that. Put that all into like a methodology file. As they go through to… I like having all that, because I can use it to validate. But the other cool thing is that, like. When I want to make adjustments to a deck or a chart or something, they don’t have to rerun everything.

Unless it actually required it, unless it was like, oh, it doesn’t have those numbers already stored, and it’s like output files. So I try and create as many checkpoints as possible. Like it’s kind of like… I don’t know, like, if you’re playing a video game and you, like, have all checkpoints along the way, and then you die, and you just go back to the checkpoint, like, that’s kind of, like, um, how I try to set it up. But I definitely hear you around, like, editing. I think other things is, like, try and set it up in ways where you can do. manual edits, so I haven’t done this in the AI analyst here, but yesterday one of our colleagues, he was like same thing.

He’s like, man, I just want to change like Claude’s giving me weird words like that one thing that was like. I don’t… what does it say? One. It was like 1 1 1 it was like some intense slide like he doesn’t want to rerun the whole deck creator. So he worked with Claude so that instead of a Pdf, it’s. all made into a Pptx that you can put in Google slides, and you can just like edit the text himself. And then you can go and manually do stuff, so… My number one answer for most questions will probably be asked a lot, and then ask Claude like in a way where it’s like, not just, like, immediately one shot.

What’s the answer, but like, hey, like, let’s go into like architect mode and like talk to a bunch of subagents and figure out different approaches for solving this. But that’s a, yeah, that’s something, um… that we really got to figure out pretty well, because otherwise it just like takes forever. So manual intervention setups or checkpoints is kind of the 2 solutions I’ve seen so far. Cool. Uh, okay, because Attendee has a good question. How do you recommend in terms of getting the data architecture in shape to be able to use this Commercially, or I guess internally, externally, whatever, like, how do you make it viable?

Yeah, so… so my recommendation is kind of like step one, do analyses you’ve already done. and see how it performed on those. And it might make some mistakes or not. And then talk to Claude around, like. hey, like, how… Is there a better way for us to store data, or is there missing data that we don’t have in tracking that could. improve on this, and like ideally, like what we’re doing at our company is like, we’re trying to build a bunch of like. more robust golden like data sets, like semantic views, so it doesn’t have to join up to a bunch of stuff.

And then we can… work into the Claude MD file or into another file within the repo, like, hey, if this kind of question comes in, just go use this like. data set we know has all this stuff, and, like, a semantic view, rather than trying to, like. piece a bunch of things together. So, like, simplify it as much as possible. I think, like asking it where gaps are is really helpful, and and you can make. You can do this very simply. You can literally like run an analysis. And then run it a few times to the point where I can get it. It can nail it exactly. And then you can say, like, can you create like a requirements doc?

Um, that I can work with my analytics engineering team and send them so they can go, like. develop a semantic view of this. I like. That’s what we’re in the process of doing. I think we’re going to learn a lot about this in like the coming months. probably the coming month. This is probably something we should, like, really dive into and can focus. And then the boot camp. But like, that’s that’s something like I don’t have a great answer for right now. But I do think that’s the next step. I think it’s like the next path is like going to be getting good data sets, because that speeds it up, and it makes it more reliable.

Yeah, I was wondering, uh, if… Data modeling patterns like OBT, or dimensional or relational, or whatnot. would align better with Cloud Call Code. Yeah, yeah, it might. I’m not sure. That’s a really good question. I think that’s going to be like. One of the next jobs, honestly, it’s not just gonna be like data engineers doing that job. I think, like. That will be something I definitely work on, where it’s like once you get, go from that, like. all this mechanics on data scientists, and you earn like thinking time. Those are the kind of problems I think will be some of the first ones probably want to spend a lot of time thinking about. That’s what we’ll definitely be doing.

a lot of questions I’m trying to see if I can boil this down a little bit. Maybe I’ll just… More questions? I’m trying to see if I have a meeting after this, or if I can stay longer. Uh, let’s see… Okay, uh… I can probably stay till 12. Attendee has a question around, um, how can we embed agentic systems You demonstrate it in Cloud Code into our existing work analysis SaaS platforms, one of the key bottlenecks I face at work is translating these capabilities into our internal analysis. products and workflows, so clients and brands can analyze their own data in the same way we’re doing it in Cloud Code. Oh. Yeah, so… you’re trying to create like an. analyst for your.

customer through your SaaS platform. I think, like… This is, like, I think this is, like, the same thing as, like, uh… People are trying to. build analyst for someone else’s domain at their work. Like. My thing I’m trying to teach here is more around like automating at yourself, where you can validate yourself. If you’re going to do it for someone else, or you’re gonna do it for your customer? You have, and you don’t know the data, like, the back of your hand, like, if you ask. a question that you think the customer is going to ask. You have to have like. feedback, so you basically have to set up an AI evaluation system where.

What’s gonna happen is if you create a system like this to answer your customer’s questions and their tools, you will think of a lot of questions that they might ask. and then you’ll give it to them. And then they’ll be like. hey, they tested this, and they got these answers right. This thing’s pretty good. Let’s ask hardware questions. And you’re not going to have tested on this harder questions. And that’s when you get into the problem of like. the non-deterministic system, because you have not. validated those outputs of those harder questions. And so you’ll have to basically set up, like, ground truth data sets, or you have to have customers who are willing to, like.

like, give the feedback to the system itself around like, is the question wrong or not? So like, this is what a lot of the like. You know, agentic analytics Bi tools are doing right now, where, like, if I’m going to use a snowflake or something like I can. I can go in and like. Try questions, and then label like if it get it got it wrong or not. But like your customer would be. Saying if it’s wrong or not, or you would be working with your customer to say what’s wrong or not. But this is a little different than like building it for yourself. it will be harder and slower. But. I don’t know, that’s a great answer, but it’s also kind of like not what I’m focused on.

generalizing is much harder than you… Next question. Yeah. do it in places where you already are an expert in, and then you can start to demonstrate, kind of, like, that value. over, I guess, building a lot of really complicated systems to try to solve every single use case. That’s generally true of the approach in AI. Right now. Okay, Attendee, I have another question. doing a lot of allergy. Yeah. Attendee has another question, how do you approach building data use cases for cloud code? As in thinking through what use cases the agents can do additionally to answering business questions or building slides?

How do you approach building use case for good thinking through which what use cases the agent can do additionally to answering business questions. So I just try and do whatever I’m gonna do in my job. Basically, when I got in person with Rabian High a few weeks ago we basically were just like. Let’s do our job. But every single thing we do, let’s see if AI can do it. I’m talking about reading my calendar events in my Gmail, reading my slacks and creating the notion docs and doing the analysis. Like, I just work with cloud code to see what it can do and like talk to it and go through that like architect kind of planning mode. and just push as much as possible.

Like, the way we kind of would work was like we sit down. We brainstorm a bunch of ideas. It’s all recorded. We save that to a transcript. That transcript talks about all the things we want to try and automate out. We give that to Claude code, and then we see what it comes with up with, and we do like a brainstorm thing with Claude code. But. I think I don’t try to go do things I’ve never done before right now. That’s not what I’m trying to do. I’m not trying to answer questions I’ve never answered before. I’m not trying to be like. solve problems that I don’t solve in my regular life.

I’m just going through the stuff that I’ve done for the last 10 years and just like seeing if like I can get AI to do it because then I can validate it really easily. And it’s like stuff I know. Well, so when it when it’s wrong, I can’t just… I don’t only have to say it’s wrong. I can say. Hey, you got this wrong. I think you… like, let’s talk about it, but you probably got it wrong because of this, and try doing it this other way. That’s… that’s where I think about use cases. I just I just kind of like. have them as what I do regularly. Cool. Uh, let’s see… Attendee has a question. Today’s demo involved fairly straightforward analysis of internal data, not to imply this is simple.

Curious about how this changes if you’re, for example, an operations researcher working with customer-provided data. For example, more data normalization, feature selection, model choices, etc. Yep. Yeah. So 2 2 things there. One, I think it works… I’m using this at my work. It works with. Customer data, so long as you need to provide it more business context and like connect to other sources, so long as you have the business context around it and can like nurture it and teach in the right direction. So, like, at our company right now, you know, we have, like, 10 data scientists or whatever.

They’re all working with customer data, but they’re all trying to do it in their own domains separately, building up the AI analyst. That’s how I would say. And like, it’s pretty complicated stuff. in terms of like, I have metrics that I’ve just like completely invented out of AI evals that aren’t created at any other company that have to do with, like. comparing certain text things to each other and like turning them into embeddings. It’s like complicated metrics that, like, it can’t just, like, look up, like, oh, what’s their funnel conversion rate? Although those are those are easier for it to learn. I have all those things already in like. repos and GitHub and pipeline. And so I feed it.

I basically clone a bunch of repos for all of my data science analyses in the product I work with, so I can give it all that context. you know, this could be stuff that’s like many tables, many CTEs, and as long as it can go through my query history. It can figure out like the patterns I use and store that in its knowledge base. But again, I have to validate it and kind of like point it in the right direction. So it can use… it can work on complex sets is my thing. Now, the other question around like feature selection or building models and stuff. So this like I’m we’re we’re staying pretty strict, just like product. analytics like segmentation, trends, root cause.

Yeah, simple things like Simpson’s paradox, like it can go do a bunch of EDA and like analytics thing. I’m not trying to get into like the data science machine learning realm yet, because then I think that gets a little more. like quite a bit more nuanced in terms of, like, why are you picking all these variables for your model? Why did you transform them these ways? And that takes me more time to validate. I think it probably gets there eventually, but… Um, we’re starting simple with data analytics, just because a lot of that is really mechanical work where you don’t have to. think too much.

Whereas, like, if I’m trying to build a causal inference model, I have to think through like what are all the variables, internal and external, that like path might whatever I’m inputting to the. the effect on the output, and I think that’s a lot more of that thinking stuff. I think for a while, like, yeah, data scientists are going to be doing more of that work than the mechanics analytics stuff. That’s… Did that answer your question, Attendee? Man, I… dude, I thought all bases had 4 strings. I didn’t know that. They went up to 8. you have to squint. Cool.

I think there’s a flavor of questions, um, from, uh, I guess, spinning off from Attendee’s question about, uh, Do employers allow folks to, I guess, use external agents on internal data and, I guess, kind of, like, more sensitive things? More around, like, organizational navigation and, uh… Making sure we can do it. I can take a stop at it, and then, uh… Yeah, so, uh, generally speaking, it is… Yeah, yeah, you do this one. The bigger the company, the harder this is, generally speaking. Like, the more departments, the more approvals you need, uh, the harder it is.

Not to say that there’s not a way to sort of, like, strategically kind of, like, knock down the barriers and stuff, Um, uh, in our company, we’re all going through the same thing, uh, probably in Sravya company, uh, she’s going through the same thing as well, where… You have to, uh, kind of, like, um, understand the dynamic of your company. So, for example, um, I have… a contact at a very big company that says they’ve been stuck at, uh, just Like, literally, kind of, like, all the red tapes for years, uh, trying to do… trying to do this, uh, because, you know, obviously they have better… they have much more strict controls around, uh, things like that.

Uh, or you might be at a startup that can just run with this, like, tomorrow. So, it ranges across the board. I think the, sort of, like, uh, you know, there’s really no one-size-fits-all sort of question, uh, sort of, sort of solutions. The thing would be kind of like… the thing that constantly would help your calls is to get people excited, like, just generally. Um, you know, demonstrate the value of it.

If it’s not possible with your internal data, perhaps showcasing it with something that is, uh, you know, like, not no risk at all, or whatever, and then sort of, like, start getting buy-ins, like, getting people to see it, getting people to understand what the possibilities are, and then they’ll become your ally and go from there. The more… the higher up that you can do this with, the faster you’re going to be able to get to where you need slash want to. So, there’s sort of like that, huh? that level of playbook that, uh, that would be very specific to each company and their culture. What else? Uh, cool. So, let’s see.

So, there was an interactive aspect of that Question… how can stakeholders verify something on the AI deck? How can they? Oh, like the output deck. So yeah, you can, uh… Like, verify the numbers, correct? When you send it to stakeholder same way you would do it right now, right? Like when I send out a deck or I send out a doc. when it comes from me or someone else. Usually I have a link to the query I ran. What I do with mine, and I didn’t do it on this one, but what I’m starting to do internally is. With my deck. It’s accompanied by Doc. Notion doc that has all the queries ran, and, like, they can go run on those themselves if they want to do it. But I would say, like.

Yeah, Claude like knows what query around like just have have it run the record the queries like usually when I report my. Decks out for the past 10 years, I’ll have, like, a link, and it can go into… Snowflake or whatever Databricks dashboard or whatever where I’m pulling that number from, you can set the same thing up here. Um… I haven’t figured out how to, like, make Claude code, like, create a new, like, worksheet in Snowflake, so it goes right to the Snowflake worksheet, and, like, they can run it. Like, that’s what I do with my analyses. So right now, I just have it storing all those into like a notion page somewhere else, and they can verify the SQL queries there.

Cool, uh, let’s see… Sorry, lots of questions here. I’m trying to… Okay, I think Attendee has been… has been diligently posting the same question throughout, so maybe we should answer it. Okay. Thank you for this demo, great questions. In terms of moving to implement in our jobs, how to connect with large data, or… data in cloud, or I guess data… data warehouses like BigQuery out to build semantic views across different areas of business. Mm-hmm. Yeah. So you want to kind of tell like just ask a lot like connections to. to your data warehouse are like what like snowflakes connection is like 6 lines of code to open a connection.

They all snowflake also has cortex, which is basically like cloud code hosted by snowflake. So you can even like go direct through there. But you can just ask Claude like. Hey, I use BigQuery like let’s get you connected. If you if you if you. Download that repo. We probably look at it in here. There should be a skill on this. So, like, it’ll actually ask you. for data connection. Yeah, connect data. Skill MD. So if you back… so these are skills. If you go to slash command connect data, it’s going to walk you through, like, oh, even has BigQuery baked in over there.

So some of it can create new connections for you, but like for BigQuery, for instance, I already put in the code snippet for how to connect to big. query, it’s gonna ask you, like, there’s probably different ways to do it, right? You could probably OAuth in through external browser, or you might it might want like a username and password that stores in your environmental variables, or use, like. a like a path. So like a token that you store in your environmental variables. But you can just. basically have Claude walk you through that process of like getting connected. Um, in terms of how to build semantic view across different areas of the business. I’m sure there’s many.

There’s many courses that of data foundations that would teach teach us how to do that like. I’m not a data engineer. So like, I’m not actually probably going to do that. We can go into a little bit of our thoughts on that in the. in the boot camp, but honestly, like. Yeah, I would say it’s like, look that up on on some stuff on YouTube or documentation. I think that’s like beyond. like, just like building AI analyst think flawed code, although Claude can definitely help you build those semantic views based off of like the query user running here. But we’re not teaching a course on how to like build all the data foundations.

at your company, we’re not going to be able to do that in 2 days, or 3 hours. So, I think, Shane, like, my question was more focused on, uh, around, you know, uh, like, because there are many different nuances to each different tables that we usually have. And as you’re working on it for, let’s say, 2-3 years, you’d usually have, you know, a good sense of all of those things. Yeah. So I think it’s about, you know, how do you give that extra context time to time? I mean, one is the thing you mentioned that you keep teaching it. But when you kind of start, I think there’s, like, a lot of stuff that we do.

Maybe, like, it’s not, like, just 5-10 tables, it’s maybe, like, 50-60 tables that we, right? So in that case, how do you kind of get started quickly with that context for that? Yeah. Uh, okay, okay. I see what you’re saying. How I did it is. Let’s see. So who’s this? Knowledge bootstrap thing. Basically, how I did it is I told it to go look at my query history for like the past 30 days to see all the tables I usually pull from, all of the joins I usually make.

all the columns I usually pull, and then I have it start storing that in a… there’s, like, a Json and YAML files in here that will, like, when you ask about certain like if it was to ask about NPS or whatever, like, we can… it just could look through the tables now. But when there’s, like, 50 or 60 the like agent would route to like first look that that knowledge base. It’s like, or this is a question around NPS. Let’s ignore these like 50 or 60. Let’s look at these other ones that are based off Shane’s query history. That’s one way I’ve done it so far. I’m sure that like.

data engineers and data platform folks are gonna have like a, uh… Much more sophisticated answer for this later on, but that was like a pretty easy way to get started. It’s just like, look at my query history and then correct it along the way. Does that answer? Is that more like your question? Uh, yeah, I think that that provides a direction. And do you… did you face any kind of… like, challenges in terms of, like, uh, like, data security or any kind of those aspects when you, let’s say, connect it with one of your actual data sets, like, not just, you know, CSV files, so any concerns that you face from, maybe, organization or just by yourself? Yes.

Yeah, yeah, you just have to, I mean, you have to work through with your security team and your legal team in terms of like what you can use and what you can’t like you might need to have like. You probably want like a 0. data retention policy with anthropic. So that’s like something that has to get worked out through your organization. Anthropic, because you don’t want them keeping your data. You might not. You probably don’t want to have stuff locally on your machine. So like, if I’m using with snowflake, I use cortex instead of Claude code, because. Then all the stuff’s running in like snowflakes. environment, not on my machine.

If I’m using Claude code to analyze stuff, I do it in like SageMaker. So it’s like in a secure hosted environment. It’s never on my laptop. So like. That’s, like, something you definitely… that will come up, um, that you want to work through your organization in terms of, like, how can I do this securely? Like Hai said before, I would say, like. Do it on non-sensitive data, or if you have to public data, whatever. locally first, if you don’t have those secure systems set up already. But, um, and then you can kind of start, like, trying to get buy-in for, like, hey, can we use… Claude code.

Can we like negotiate our contract with them to, like… Um, basically have them see your data, but, like, there’s gonna be a lot of security implications. I feel like the security role is probably gonna, like. be become very important, and probably a lot of people you’ll see AI security roles jump up in the next year or something as people try and figure out how do we get this stuff unlocked in a safe way and not bottlenecked by this one org. But definitely. Yeah, as I mentioned, like, that’ll be a bit of a challenge. Thank you so much once again. No problem. Good questions. Uh, maybe we… so, let’s see… Maybe this question, and it’s probably around time that we can wrap?

Yeah, I need to stop before my next meeting. Okay, cool. So, how do you interpret the recent rise in job titles, like AI data scientists, AI data analyst, and AI product manager? I know it likely depends on the company’s level of digital transformation, but do these titles usually refer to roles that genuinely require people to work deeply with tools like Claude and other LLM systems, or… Are some companies adding AI to traditional roads just to attract candidates and ride the current trend? Oh, I’m sure it’s both. I’m sure there’s probably more people doing the latter than the former, because this wasn’t working very well before a few weeks ago.

I think it was… it’s been working for product management for a lot longer. And so I think there’s generally a lot of people who are. product managers, and we’ve interviewed a bunch of AI product managers on our podcast, and they can mean different things. They can mean you are a product manager building AI products. It could mean you are leveraging a bunch of AI tools. to do your product management, like we talked about here at Cloud Code. And there’s a third one that I can’t remember what it means. But there is a YouTube episode where we talk to a mom Khan just about this so you can check that out. I think so that’s I think those are real rules, I think.

AI data analysts and AI data scientists, probably same thing. There’s probably some people who are like. I… like, I do for me, like, I do… I’m a data scientist, and if you look at my titles, like. Data scientist, comma, AI evaluations, because I don’t do as much product analytics work as I used to do. I do more of like AI evaluation work. for like AI product that we’re building for legal tech. Um, which is a very different. It leverages a lot of foundational skills, but it’s a very different job. I think like someone could easily call themselves AI data scientist if they’re doing that. And then I think something like this where you’re leveraging it a lot in your workflow.

Sure, I could see that coming up, but. I don’t… I don’t know. What do you think, Kai? I think… A lot of it’s probably, like, people just putting it on there, and… but I do think it’s a real role that will happen. Like, I think there will be a division between. people who use AI to automate out a bunch of their workflow into a very different job than people who don’t. And I think people are going to want to make that distinction. Yeah, Attendee actually has a really good, uh, analogy, which is, uh, personal view as a PM leader, AI product manager is like calling water wet, or soon will be in any case. I think that’s exactly right, at least my point of view is most companies Yes.

a lot of… like, we’re really early right now, but then the advantage is gonna get wiped out pretty quickly. the companies are generally looking for people that know how to do this kind of stuff. But then… so they’re trying to attract talents, just like that, but they don’t know how to do it themselves, or what that could look like. And so, you know, like, part of that could be… Uh, hey, someone can come in and help us to do this, but then they actually don’t know what the requirements are. Um, but, like, that’s… that’s gonna change pretty rapidly. Um, if we sort of, like, embrace that quickly. that’s where… that’s how it benefits, sort of, like, all of us. Yep. Do agree.

One thing Shane and Hi, a few people have been messaging me personally, and also I see that here. How is what we shared so far different from the boot camp? So what will be the paid bootcamp look like when compared to the free one? If you could maybe want to take it over? I could add after, and I can. Well, first of all, there’s a lot of questions to be answered that we’re just not going to answer here. So I will definitely answer those there. But we’re going to start from scratch. So we’re going to like go deeper into how to create all of these different.

files and and Claude, how one works with another like how the Md file maps to skills and engines, how not to like create one that’s like going to get you like. I guess, over engineer one, how to make it fit right to your company. We’ll walk through with you, like, your own data set to like get you set up by the end of the weekend so something’s running on your own. Like, honestly, like. go take the free repo down and. If you work with that and can. Bring it over to your own company, port it over, and it works pretty well, like, that’s awesome. Like, maybe you don’t need to take the boot camp then, but if you want to, like, learn like the process we went through to, like.

build the full agentic system, like, step by step. That’s what that’s going to be about. I don’t know, like we can share a… the outline of the course, too, maybe is, like, more helpful. Hi. Yeah. Yeah, definitely. Do you have anything to add? One thing I could add totally with what Shane said. So when when Shane started doing this, and he was like, my mind’s blown, and I started looking at it. There was a ramp up that I needed to understand because my day job is crazily busy, and I don’t get as much time to work on the most exciting stuff. So when Shane shared with all of this, there was a ramp up time for me, a learning curve for me to get to a place where we got here.

So all of that learning curve, understanding why the structure, why we came up with the skeleton in the first place would help you. take this and do anything with it. Someone were asking about product manager, someone were asking about how do you do this for finance analysis? Like, you know, things like that. So when you know and understand the skeleton and get to the details of it, that’ll help you take this and use it anywhere. And that’s what we would help you with the boot camp, the paid boot camp.

Yes, like Shane said, we shared a lot of value here already, and you could get most of it yourself when you look at the GitHub repo, but taking that context, understanding the whys and what’s behind it to replace. your existing workflows at work, that could be the paid bootcamp, yeah. trying to pull up the outline right now. Yeah, and the concepts there is pretty extensible, like, the good thing about AI is once you understand the… uh, like, the foundations of it, you can basically… built very similar things for, for example, like, uh, like an AI PM, or like, uh, whatever PM, or sorry, AI whatever, like, AI any title, if you’re the domain expert in that area.

So, it really is Just like, uh, we start from analytics, because we’re, like, you know, we’re really good at it, and so, you know, like, but that sort of thinking extends to everywhere else, basically. And for analytics, if you want to get really good at it, we explain the concepts as well. That’s another longer course, the five-week course, which is AI analytics for builders. So you could also enroll into that. So two ones. One is a paid bootcamp, which gets into the depths of understanding Claude Code and the setup and the skeleton and all of that. And how do you use it for multiple reasons?

The other one is if you yourself want to understand the analytics in depth and the metrics, the experimentation, and, like, all of that, like what we do in data science in product analytics world, that’s another course for you. Yeah. Yeah, I’d say the 1st hour. of the boot camp is gonna be, like, pretty similar to what we did here. And then, you know, we’ll talk about the problem. We’ll give you a demo. And then after that, we’re gonna. set up everyone with Cloud Code, like help them like know what, how to connect to all their repos. We’re going to. build… get you to build your first agent and skill. We’re going to iterate on those agents and skills.

We’re going to analyze some external public data sets. We’re going to start connecting to your data sets at your company. We’re going to teach you how to like build like the kind of like learning. feedback loops that’s like we’re specific for your company. That’s kind of like the high level, I would say. And we’re also just there to answer a lot more questions, I think, too. Right. With that, though, I need to I need to use the restroom before my next meeting. But great questions, everyone. Thank you, everyone, for staying so long. And yeah, if you have any questions, just DM any of us on on LinkedIn. And hopefully we’ll see you at some of our future sessions.

You can check that out at AIAnalslab.ai. Other than that, have a great weekend. We’ll send the recording out, too. Thank you, everybody.