The MCP Sessions - Vol 2: Supply Chain Analytics

0:01All right, let's go. Alex, how you doing? >> Doing great. Welcome everybody. We're chatting MCP today. So, uh, welcome to the Motheruck channel. Uh, and, uh, we're happy you're here. So, I want to talk today a bit about what does it feel like to use AI in the data world and really think about how can we take

0:21advantage of an MCP to really play around with our data and get some deep understanding of it without writing 10,000 SQL queries all by hand all ourselves. Uh so, um first we'll we'll say a quick hi. Howdy. I'm Alex. Uh I'm a recovering data scientist. I spent uh nine years at Intel uh in the supply chain world. So, I'm really excited that

0:41today we're focused on a supply chain data set. Uh, so really definitely near and dear to my heart. I've stared at many millions of rows of this uh shape and size uh uh before and so happy to to look at it with AI today. And then here at Motherduck, I'm a developer advocate and um you bet. I'll pass it over to you

0:58Jacob. >> Amazing. Thanks, Alex. Yes. Uh I'm super excited to look at this too as someone who uh also worked a little bit on supply chain. Uh, I ran uh the business systems team uh at uh at a toy company

1:12that some of you may be aware of called Funko um in the run-up to going public which is very fun very fun journey. Um and so I have lots of time lots of fun with tracking things like on time info and other supply chain goodness and doing all the optimization linear algebra to make that all work. Um, so

1:32I'm excited to see how much of this we can defer to AI. And um, Alex, why don't you show us how easy it is to kind of get this all configured so we can start using it.

1:42 >> You bet. All right. >> I'll throw you on the stage here. Is that right?

1:46 >> Wonderful. >> You bet. So today we're we're using uh Claude, although we'll point you to our docs if you're using another um another AI tool. So um the motheruck MCP is uh

1:58remotely hosted. So there is no local setup that you need to do at all. Uh it's just operating on the motheruck side and we're we're hosting it on your behalf. So to add it is as easy as going into your cloud settings and you're headed to connectors.

2:14Once you get to connectors uh we are in the cloud marketplace. So if you go to browse connectors and search for data warehouse you will find motherduck and all you need to do is add that in.

2:28 >> Yes. and you will have the MCP fully operational. So MCP stands for model context protocol. It's a way to have a a web endpoint to where your AI can talk to it and and gain skills that >> aren't easy aren't easy to do for an LLM. And so this server on our side that we're hosting for you allows you to get

2:48insights directly about your data set. Um because LLM's by default don't know anything about your your your data. uh especially your private data. If it couldn't have been trained on it in the public internet uh it wouldn't it would have no knowledge but now we can uh we can give that knowledge to the model.

3:06 >> So that's all you need to get started. I did already set that up on my side. So it really is you know couple clicks there and you're you're good to go. Um, if you wanted to do this with a different setup, if you head to our docs and go to our SQL reference and our MCP, you'll be able to go ahead and set it up

3:22with a variety of different, uh, tools, cloud, cloud code, chat, GBT, cursor, as well as others. So, uh, MCP is an industry standard now, uh, which is great. So, you can use it with whichever AI tool you'd like. And if you're here, follow along. Uh, see if you can get it set up and get it going. Uh, uh, we've

3:40got a free Mother Duck account. you can hop right in or a free trial and then you could hook this up pretty quick.

3:48 >> Sweet. So, wanted to quickly touch on what is our data set and where we're getting it from, but then very quickly we're going to go explore it with MCP.

3:57So, to load this data set, this is the Modok UI. Um, I just did a couple of SQL statements. So this is our SQL notebook interface and uh I created a database to to hold this data and I said hey let's focus on that database and I created a table called shipments and that's what we're going to be

4:15exploring. It's a single table of data set and I got it from hugging face and you can see it's as easy as selecting from a URL that ends inpark here and we can just go out to where hugging face is storing your data or public data and uh grab it and load it into mother and that works for all kinds of different spots

4:34where your data can live. It can be S3, Google Cloud, Azure or any file in GitHub lot of different things. And then uh from there I'm just selecting that data and I'll just run it. takes about 10 seconds or so to to pull that in. And it's uh about 180,000 rows and we will explore it a lot more uh with AI. You

4:54can see though that within the motheruck UI, we do have some exploratory capabilities here with the uh column explorer. We're going to ask the AI to go go a couple levels deeper if we can.

5:04Um if you're curious about how do you get a data set from hugging face, this is the link to this particular data set.

5:10Um you can uh they automatically generate parquet files for every data set on hugging face. So you can click on the uh megabyte indicator of how big your files are. You can navigate to where the file is and then they have this handy dandy you know download icon that is very obscure but but you can uh you can grab

5:31the the download here uh from the download button and that will let you just get that handy dandy URL to to put right into a SQL statement.

5:42So this data set, let's uh let's let's see what it is. So how how easy is it to figure out what what's in this data set?

5:50Let's say please give me an overview of the shipments data set. Let's do it.

6:04All right, we're using Opus 4.5 as one does. So now we have the first step here and you might have an authentication step here. I'm already logged into motherduck um but it might prompt you to say hey please log in. You can log into motherduck through you know github credentials, Google credentials or username password >> and you're you're connected. So so what

6:25is happening here? The first thing is that it looked in the catalog and it said, "Hey, is there anything in the catalog about shipments?" And it got back an answer that hey, there is one table called shipments and this is exactly where to go to find it. So found it first try.

6:42It's it's it understands what I mean when I say shipments. It connects it to my actual data and then from there it says, "Hey, what's inside?" And it'll figure out all the different columns that are there as well as the different data types that are there.

6:55Yeah, I think it's worth mentioning here, Alex, that like we've built tools that are designed for this type of workflow, right? So, it hasn't written any SQL in those first two calls, right?

7:04Those are kind of API endpoints where it can take a, you know, the search catalog just takes a generic term says, hey, I need to find where shipments exists in your entire catalog. And so, it does that in a very constrained manner as instead of saying just like, hey, let's let's look through the entire, you know, entire database in a very naive way. It

7:21just says here's a tool that will do this for you, which is very, very helpful.

7:25You bet. And you saw it converged really really quickly. >> Yeah. I mean that was basically founded immediately which is great.

7:32 >> Yep. >> Absolutely. >> And then from from there is where it starts writing some SQL for us.

7:39 >> And you know as someone who's been a SQL jockey for for many years and considers myself one still this is pretty crazy that it is uh that is able to just start with a table and start writing a lot of queries on my behalf. and and you can see what queries it's generating, which really puts me at ease as a SQL person.

7:57I can actually see exactly what's happening. So, this is not a black box. This is a completely glass see-through box where it tells you exactly what SQL is running. Uh so, you can proofread it or or understand it for your own use as well.

8:11 >> Um, so what's it look like? What did what did AI uh what did what did Claude do when I asked it to just give me an overview?

8:19So, it tells us how many order line items we have. So, it already knows that it's it's a a line item specific table.

8:26We've got 66,000 orders, 21,000 customers and 118 products, and it looks like it's uh almost two years of data from 2015 to 2017. And it's even able to detect kind of the the domain of the data that is e-commerce and it's supply chain oriented. Um, which is pretty great. And it picked up that some of the columns are related to

8:49profitability, which I think we can explore as we go. >> Key findings, as Jacob was talking about with on-time delivery, the first thing it found is that 55% of shipments are late.

9:02That's pretty neat that it was able to look at a data set of shipments, automatically decide to look at things that were late or not, and then tell me how much is late. And 55% late strikes fear in my heart. Like, I feel that deeply. 55% late like oh I feel it. So so thanks for for for looking at that

9:19for me. [laughter] >> Now it's going to go even deeper though. >> It varies very dramatically by shipping mode. If you do first class you're you're going to be you got a 95% uh risk

9:34of being late whereas standard is 38%.

9:38So it is actually jumping to a very interesting conclusion that says we are being very optimistic about what we can achieve on first class shipment. It says hey seems like you can't actually deliver it as fast as you are committing which is which is very interesting.

9:55It looks at our different markets of of what we're looking at um with Europe and and Latin America leading in volume.

10:04Late delivery risk is remarkably consistent. This is kind of neat. I mean, how many queries did it run on my behalf? Like what, like six or seven?

10:13Eight different queries. You can tell this is a lot of layers of detail in just an overview, right? This had to figure out late delivery was a problem.

10:21Then it had to go say, okay, regions, that's probably of interest. Now, how does that metric of interest relate to that category of interest? It also breaks down our products. We got fishing, we got cleats, we got camping and hiking.

10:35And [clears throat] notably, computers has the highest profit per order, but very low volume that tracks. Makes sense.

10:45We got our consumer segment. Fantastic. So, there's there's a lot of meat here. And this is just our very first thing. And I asked it, hey, what should I look into next? And so, it gave me a couple of ideas and we can look at these or decide direction. Um, [cough and clears throat] the delay delivery paradox. Why are

11:05premium shipping modes late all the time? It's very interesting. Uh, payment method transfer payments show notably lower late delivery risk.

11:17Interesting. Now, this this is a shockingly interesting question here. Is this a real operational effect or a customer segment correlation?

11:30Claude is asking if it's correlation or causation. As a as a stats fan, that's pretty pretty sweet. Uh I'm I'm impressed. Um

11:40mentions fraud is something we could look into. We can look at profitability drivers and then we can also do a deep dive on geography.

11:50Well, I think it'd be interesting to do a deep dive on late delivery a little bit and then maybe >> agree. Uh, so what do you think, Jacob? How would you how would you dig into the next level of detail on the late delivery side?

12:05 >> My suspicion, okay, I have not looked at this data set yet, but my suspicion is that if your first class deliveries are always late is probably that like you're upgrading people to first class once you get late already. So um I would test

12:21that hypothesis being like is there anything you know ask like is there anything deeper around the first class deliveries like you know um what what other patterns are in there. I think some of this also gets into like operational context right which is like I've definitely worked at companies where on the supply chain side um if product wasn't going to get into our

12:40warehouse on time we're like okay that one can go air instead of you know freight or whatever. Um, and so, uh, I

12:49would be cur my my first instinct is actually that like people are getting upgraded to it and it's like trying to save late delivery that might be within SLA on a short delivery but like it's or on a longer delivery but is not getting updated. Right. So,

13:06 >> I'm going to ask what correlates with late deliveries as well just to kind of seed it a little bit, you know, with with what >> Yeah. what double could be related.

13:17So let's see what it's what it's looking at first. First things first, it's looking at our lateness year and late percentage and it's grouping by shipment mode and then days for shipment. So sort of what we committed to and then and then the reality.

13:33This is very revealing. Interesting. Well, can't wait to hear what you're thinking here, Claude. [laughter] One thing I've noticed, I don't know, uh, Alex, I don't, one thing I've noticed is that Claude is very curious, which makes it a very nice analyst partner. Um, not all models are this curious. So you, you know, you don't have to prompt it too much for it to

13:51kind of go look for something, but on the other hand, it means it can go a little bit off the rails in terms of the analysis. So it's a balance for sure.

13:59 >> Oh, that's uh, you bet. That's explore versus exploit. Uh, >> yeah. [clears throat] Yes, exactly.

14:04Totally totally totally right. >> Love it. So something that is I think very interesting right is is as someone who's written a lot of SQL and used a lot of BI tools part of the value of a BI tool is to really dig in and ask a lot of questions really really fast.

14:22 >> Yeah. >> You can see how many queries are being run here as well and and we're able to ask a lot of questions very very quickly much faster than we can write the SQL as well. But it's it's very um open-ended.

14:34We're not building a dashboard that's, you know, where we're linearly exploring something or we are having to use some sort of pre-built asset to do our exploration. This is from scratch. Um, where we can we can just kind of explore in all directions at the same time. And I can write SQL pretty fast. I can't write it this fast.

14:54But um, no matter how clicketity clacky my keyboard is, which is very clickity, very clacky.

15:03 >> Totally. Let's see. So, as that is thinking about our follow-up question around delivery, I am very interested in the profitability piece. Something that comes to mind uh from my days in supply chain is sometimes you are very sure that the expensive things are delivered on time [snorts] and the cheaper things it's okay.

15:23 >> Oh, interesting. Yeah. Okay. >> There's a relationship between probability and delivery. >> Uh just to just to get a sense. Yeah.

15:33Yeah. >> And then maybe anything else that correlates with profit. >> Oh, that's a good that's a good that's a good thing to look at. I love that. I love that thought. That's a good thought.

15:40 >> I just a lot of pain. There's a lot of pain I think [laughter] behind that that insight.

15:46 >> I mean, you know, trial by fire. Uh, you know, [laughter] >> okay. So, so Cloud has run something around 20 different queries on our behalf. Yeah. 21 queries. Now let's see what what insight we can gain from our data after one or two minutes of wide openen exploration.

16:05So this is this is uh clause defining the core problem as being unrealistic. So if we look at first class shipping mode, it is 100% late.

16:17The 5% that are showing as on time for first class are actually cancelled orders. [laughter] Wow. Okay. Hey, so we found an interesting nugget in our data set.

16:27 >> Oh, that's amazing. That's >> It said the the only time it was on time is if it's canceled, but that's not really what we want, right? But but the the model was able to detect that's probably not what we want either, which is very impressive.

16:41Same day, pretty pretty rough rough time. Second class is actually even worse. Very interesting. Um let's see what else we found. First class is structurally broken. Yes, that that sounds about right.

16:55 >> We we do see that. >> Yes, agreed that actual shipping times are remarkably consistent.

17:02The actual days don't vary by market, geography, product category, order size, customer segment, day of the week, or season.

17:08 >> Wow, >> that's very like like someone's plugging that number or something. >> Seems like operationally things might be humming along.

17:18 >> Yeah. >> And maybe we're committing to different levels for for various business reasons.

17:22 >> Yeah. Yeah. It's even suggesting that the the data might be flawed, saying actual data might be rounded instead of real real data.

17:34What else doesn't correlate? Payment method makes sense. Customer segment. Oh, interesting. Order values do not correlate.

17:44So, so we'll see if profitability correlates, but >> Oh, interesting. Yeah, because you asked that question about profitability earlier. Yeah.

17:49 >> Right. This is the value of the order, but not necessarily the profitability. >> Not necessarily, but hopefully they're sort of tied together.

17:55 >> Probably. Probably. Yes. Oh, it already checked days of the week or month. Interesting.

18:04Transfer payments. All cancellations are transfer payments. Okay. Interesting. Found that those are correlated. There's some geographic variation.

18:18All right. I wanted to explore if a little bit more about the weekends and weekdays because that's a very common shipping thing, right? Is is if something goes out to ship on Friday, it might not get there till Monday. So, I want to see if there's a correlation between what day of the week something ships on and to whether or not it's

18:37delayed.

18:43Feel

18:56like it probably took me about as long to type that out as it's going to take Claude to answer my question. [laughter] >> Wild times.

19:03 >> Yeah. Oh, yeah. I mean, going down these little side quests, um, this is such a nice way to just explore and understand what's what the heck's happening here.

19:13 >> Oh, and look at this beautiful thing. I did not have to memorize the strip time syntax. It figured it out. It's extracting the daily. Oh, man. I just I won't miss that a bit. I won't miss that.

19:27 >> Not at all. >> What other fun SQL syntax do I get to to not have to deal with? Yeah. Very very fun. Okay, there's a small but real correlation. Although it varies by shipping mode overall. Very weak.

19:421% different. That's pretty weak. Yeah.

19:47However, first class >> indeed. >> It's a stronger one if it's a shorter delivery. That checks out logically.

19:57Like if I need to get it there in a day, the weekend's going to impact me more.

20:01 >> Yep. Yep. Yep. Yep. That's pretty sweet.

20:05Well, awesome. Well, well, that was a couple things I was looking at. We can go down the profitability angle. Uh, what else?

20:12 >> Yeah, I think I would ask I I think you're right. Like there's um we Yeah. How Yeah. Yeah. Just just ask about Exactly. Yes.

20:32It's coming along. Yeah, very very interesting.

20:44I like that it is able to explore so many different uh directions at the same time and then tell me what is not important really fast.

20:51 >> Yeah, get to know fast, right? It's it's looking at so many facets at once. Um,

20:59yeah, that's really interesting.

21:04Let's go, Claude. Yeah, let's let's see what what is it running about? What about product level product profitability? Interesting. So, it it's asking really the exact question.

21:17I like that it puts the I like that it puts the thinking the thinking tokens into the comments >> in the sequel so you can like see what it's doing.

21:26 >> Yeah, >> I like that. >> That's great. That's a nice little affordance. I wish I did that.

21:31[laughter] >> Oh man, comments in SQL are worth their weight in gold, but they're never there.

21:38 >> No, no, no. >> What's this CTE for? You know, it was important two years ago.

21:45 >> Oh man. >> Yep. Oh yeah, there was some weird thing in our ERP. It didn't mark the flag right.

21:52 >> Did that bug get fixed? >> That's right. [laughter] >> Okay. Late deliveries have slightly lower margins. That's what we would expect.

22:04So profit margin. Oh, it's pretty close.

22:09Interesting. The profit. >> No, it's not. Yeah, it's it's all in the right direction, but it's very minor.

22:17Interesting.

22:21Margin gap is real but small.

22:27Although, when you add it up, that's 83,000 in missing profit if you think about it in reverse. So,

22:36 >> let's see. Late orders have higher margins in some of these categories. Oh, it's giving me some other explanations. Product mix, shipping mode, or it might be the data is bad. Always good to leave that as an option on the table. Uh, question everything.

23:02 >> No kidding. >> Got it. Well, where should we where should we explore next to to see what uh what we can do here?

23:10 >> I think um ask it if we're missing anything. Like are we missing anything like else that we should look at? Like I feel like we haven't really discovered anything super compelling about it, but there's got to be something in there.

23:30I like that. And that makes me think of other uses for AI as well, like PR review or bouncing ideas back and office back and forth is, you know, what am I missing? Uh it's an interesting uh way to tap into that.

23:47 >> Yep.

23:58You know, and there's also some latitude longitude. I wonder if we could see if if like the distance traveled is correlated.

24:09 >> That's interesting. Yeah, we could maybe definitely do like a spatial analysis. >> Yeah.

24:13 >> Um I'm less I'm less adept at spatial. So, this is where I really need to lean on the LLM.

24:19Yes, special is a whole sub field. It's wild.

24:29Let me just look at the column names one more time and just see if we have latitude and longitude for both things.

24:39So, we have

24:45There's latitude and longitude, but I'm not sure if that's the origin or the destination. I think that might be the destination.

24:54 >> Don't think we have quite enough to look at total distance traveled.

25:05Oh, interesting. Let's see. So, it did run a query with an error. And let's see if it was able to figure it out.

25:12It doesn't you can't drill on those. You can only mouse over to see the Sure.

25:16 >> Oh, must appear group by clause. [laughter] >> H I also feel that very deeply >> indeed.

25:24 >> I don't know if it we it should have used group by all. >> I love that we just sent it off on like a 20 a 20 query chain like with the simple with a pretty simple ask.

25:35 >> Yeah, >> you mentioned CL is curious. You're not wrong. Is very [laughter] curious.

25:43But that's much better than saying like, "Nope, everything looks great. Get this." I I love this.

25:48 >> Okay, let's see. Key findings.

25:55 >> Okay, so each payment type has specific statuses. Okay, interesting.

26:05This is saying it's it seems like it could be synthetic. Interesting. H >> all fraud is transfer.

26:13Interesting. Okay. 90% of orders we lose money on. That is a worth asking questions about.

26:25And then there's a severe section. Less than 100% negative margin. That's a big negative number.

26:36 >> Yeah. losing $300 per order. >> All customers are in the US or Puerto Rico, but orders shipped to 100 countries.

26:46 >> What's going on here? [laughter] >> That's a US-based retailer with global shipping. >> Okay, there's more to look at there. Uh, >> it seems like the company might be based in the US or Puerto Rico and then um then the customer base is is worldwide.

27:03What kind of company has 30% of customers in Puerto Rico? I want to like there's got to be a very specific

27:12 >> this is assuming this is a real data set that narrows down to like a very specific set of companies. I think >> they must operate out of you know Miami or Fort Lauderdale you know.

27:20 >> Yeah. Yeah. Yeah. Exactly. >> Right down there. >> Exactly. >> Totally. >> First order experience doesn't predict retention. Now that is very interesting.

27:31Customers whose first order was late show nearly identical retention to those with on-time first orders.

27:37 >> Wow, >> that is a very novel question to ask. Does a late shipment lead to a customer churn?

27:45 >> What a great question that I wish I had asked about this data set. [laughter] >> Very interesting. We're safe. But that's pretty cool.

27:55 >> No meaningful time trends, so we're not getting any better or any worse. >> No.

28:00 >> Interesting. Okay. Discounts don't drive losses. >> Interesting. Interesting. >> And then the data structure itself.

28:10Okay. I'm not as worried about that because payment orders to order statuses, >> I can look at that and logically think, okay, that's probably fine. Um, so I'm glad that it always gives me a couple of choices of what to investigate further.

28:21Um, as a person who's done a lot of data analysis, a lot of digging in a BI tool for insights, a lot of building content for a BI tool for insights.

28:31 >> Yeah. Yeah. >> This is this is pretty great. >> Yeah, it's a good starting point for sure. Do you want to Do we want to Do we want to see if we can get a nice visual out of this?

28:41 >> I would love that. Yes. >> Okay. All right. Just uh let's see if we can do like take this narrative and throw it into MV or something. That's what I would do.

28:51I'm just going to link this for everyone who's following. >> It's a library I'm working on that is a chart library that's built just for using inside of Claude.

29:04 >> I have to do this one though. >> No, no, don't don't do don't do that.

29:07Actually, uh I' I've refined the prompt a little bit. It's it's you don't need to do that anymore. You don't need to tell it's it's Edward Tuy. Actually, it's interesting. I'll talk about it while it looks at the skill. Um,

29:21[snorts] Tuy basically assumes that the user generating the uh, insight has all of the business context and all of his principles are built around that notion.

29:33However, when you're using an LLM, the LM does not generally have your business context. It has some of it if you've given given it to it with markdown or maybe some context generation.

29:43So telling it that it's an assistant that's designing to poke at the edges actually has led to a lot better analysis than saying you're tuy. You should still use the tuy visualization principles of course but it be it you want it to be a little bit less certain in the way that it's like framing out all of these pieces. Does that make

30:00sense Alex? >> I think that tracks because I don't know if tuftd is as much used in exploration as in like deep understanding of something. Exactly. Yeah. Exactly. Cuz like so like once you narrow once you found the insight it's like dude be like you're toughy like let's nail this narrative and structure right. But before I found the insight I'm just like

30:22poking around and like you know having it drive me towards you have a problem with late orders. Here's what to do about it is kind of like okay like relax like I can't like I'm missing some context here. Like let's keep exploring before we start you know getting to action. But um it's really interesting.

30:40Okay, so it's building it. Um sweet. What do we got here? Sweet. What do we have?

30:48 >> I like it. All right. The font is a little small. So, let me >> I don't think it'll get bigger. It's kind of designed to be Yeah. There you go. There you go.

30:58 >> Oh, look. It's dynamic. It got a little Okay. >> Yeah, that's nice. Anyways, [laughter] >> okay. This is primarily scheduling problem. Cool.

31:0819% of orders are lossmaking. >> Hey, we got a comment from YouTube. Can you make it a little bit bigger, Alex?

31:15 >> Yes. >> Thank you. >> Go one more. >> All right. >> Okay. Uh, that's probably about as good as we can go. I think >> I can maybe do Yeah, something like that. Okay, great.

31:30Total revenue, total shipments, those are great high level metrics. Makes sense. It does the late delivery breakdown by the various scheduled days.

31:37How first class is having a rough time >> and conditional formatting for us. How nice of it.

31:42 >> That is very nice. >> Wow, that's a lot of 100% late. Look at this. [laughter] Having a rough time. Oh man, that's hard.

31:51 >> Yeah. Yeah. Yeah. >> Okay. Late rate by shipping mode. Similar similar data there. Actual shipping days.

31:59 >> Oh, okay. I think we know where this data set came from. [laughter] >> It seems like those are the choices and maybe it got a little bit uh split apart there of uh >> All right. All right. Well, that's that's an interesting coail there. Um what doesn't affect it? Our market, our segment, order value.

32:23I like that it broke it down. It just told me in the text. Oh, there's >> Yeah, it bucketed it. That's really nice.

32:27 >> Bucketed it. That's That's very clean. Oh, here's the days of the week. It's not a huge difference. Uh, Tuesdays, what's great days on Tuesday, Saturday?

32:40 >> Can you give it a note and just say just say um just say uh copy that chart name.

32:45Do me a favor. We're I'm going to show you a little trick. >> This part here? [clears throat] >> Yeah. And then just say uh change the y minimum value to be more in line with this chart.

33:03Yeah, exactly. Great. Um, >> what's really cool about this is that should be like a two-line change and it should just pop our pop our analysis right in.

33:13 >> Wow, look at that. Just a couple lines.

33:19 >> Okay, read. >> There you go. You got it. >> Wow. >> There we go. There we go.

33:27 >> That is slick. I like that. [laughter] >> Yeah. Uh I've kind of designed it around the notion of like kind of you're standing over the shoulder of an analyst once you get to this stage and you're like, "All right, change that thing." Right.

33:40 >> Anyways, >> love that. Very cool. Got our margin. We have a little bit of a benefit there. Oh, we got a couple floating metrics around.

33:50 >> Yeah. Yeah, I know what's going on there. I'll fix the bug. That's fine.

33:53 >> All right. >> Little bug to fix. Oh, that's all right. This is all exploratory.

33:59Okay, it's breaking down by category. Cleat and cardio equipment. >> Interesting. Fraud. We think transfers are all bad, huh? Bad news on the transfers.

34:14 >> Okay. Recommendations

34:18under what is it? Uh, under promise, overd deliver. >> Yeah, exactly. Exactly. It's really hard in shipping logistics. It is hard. Um >> yeah.

34:29 >> Okay. Understand what's up with our profit ratio? Why are we losing 275% on some stuff? O I love the idea of looking at that a little bit more.

34:37 >> Yeah, totally. >> Nice. Well, that's a that's a pretty clean report and that's a lot of different queries that were used to build that.

34:46 >> Yeah, totally. >> I didn't have to write any of that SQL. That is wild. That's wild. [laughter] >> That is wild. It is wild. Um, yeah, this is really cool. Um, well, we're a little over time here, Alex. Um, why don't we wrap up here? Anything anything you want to plug uh coming up?

35:06 >> Let's see. I would just stay tuned with with what we're we're able to to look at here. I think if if AI and data have felt like they haven't over overlapped to you, this was the unlock for me to see how they relate together where where suddenly me as someone who like is a SQL expert, I now have a real reason to tap

35:26into AI uh to to help me whereas I didn't before. Uh so let us know your feedback. We're we're iterating on this MCP very very quickly. Uh and uh let us

35:37know what you're able to to figure out. >> Amazing. Um, I linked a couple things that I will share um that were linked earlier, but I'll just highlight. One of them is we've got an event that's I'll be doing a webinar on preparing your data sets for uh AI readiness. Um, talking about what kind of things you

35:54can add to it to improve the the uh improve the results you're getting from LLMs. Um, and we've got another MCP session this time next week for software developers. So if you're doing software development, we'd love to see uh love to show you what's possible by integrating MCP for data analysis into your uh workflow.

36:13That's all I got. All right. Uh thanks everybody for joining us. We're going to wrap up and uh Alex, thanks for finding a fun data set and we had a fun time exploring it. So uh we will chat with everybody later.

36:26 >> Yeah. Crack on and prosper. >> All right. Bye.