Building an Enterprise AI Platform in 120 Days: The Alpura Case Study

Summary

Alpura, a major Mexican dairy cooperative operating for 53 years and owning its own cow farms, migrated from seven disconnected legacy platforms — including Snowflake — to a unified Databricks Lakehouse in 120 days using Medallion Architecture (Bronze, Silver, Gold) and Unity Catalog.
The implementation established a governance layer through data stewardship, a semantic layer for consistent business logic, and guardrails for application development, and then achieved 15–20x productivity gains when building AI applications with Genie agents on top of that governed foundation.
Production applications include an Ambassador Program app monitoring store compliance and retail execution, and a Traceability app that traces dairy products from source farms through to stores — both demonstrating the AI/BI factory approach enabled by the governed Databricks platform.

Building an Enterprise AI Platform in 120 Days: The Alpura Case Study

Alpura, a leading Mexican dairy company, faced fragmented data across seven legacy platforms with weak governance and no scalable AI foundation. In 120 days, they rebuilt their entire data infrastructure on Databricks using the Medallion Architecture (Bronze, Silver, Gold) and Unity Catalog, creating a governed, unified lakehouse foundation.
this video covers the practical blueprint for building an enterprise AI/BI factory: establishing governance through data stewardship, implementing a semantic layer for consistent business logic, and applying guardrails to accelerate application development. You'll see live demos of production applications built with Genie agents, an Ambassador Program app for retail execution and a Traceability app that traces products to source farms. Attendees learn how Alpura achieved 15-20x productivity gains.
🤝

Chapters

FAQs

How did Alpura consolidate seven data platforms onto Databricks in 120 days?

Alpura used a phased approach starting with Medallion Architecture and Unity Catalog to create a governed lakehouse foundation, then layered a semantic layer and data stewardship program on top to enable consistent business logic. The 120-day timeline covered migration from all seven previous systems, including Snowflake, before the team moved into AI application development.

What were the main data challenges Alpura faced before moving to Databricks?

Alpura had more than seven disconnected platforms with no data contracts, different BI tools producing inconsistent results, weekly reconciliation processes consuming 40 or more hours, and no single golden source of truth. This fragmentation made it impossible to build a scalable AI foundation or reliably answer business questions.

What is the Ambassador Program app that Alpura built on Databricks?

The Ambassador Program app is a production application built on the Databricks Data and AI platform to monitor store compliance and retail execution performance for Alpura's field teams. It was built using Genie agents on top of the governed semantic layer established during the 120-day foundation buildout.

What productivity gains has Alpura achieved with AI applications on Databricks?

Alpura reports 15–20x productivity gains when building AI applications using Genie agents on their governed Databricks foundation. These gains come from the combination of a unified semantic layer, governed data products, and the ability to rapidly deploy new applications without rebuilding data access logic from scratch each time.

Full transcript

[00:08] Good afternoon. I know we're after lunch, so we'll probably get you excited a little bit. Um Ready for a great show? No, we will see. So, you're here sitting down um from zero to 100 in from zero to 100 in 120 days. This is
[00:24] the Alpura story. This is a real story and hopefully you'll get some pragmatic results that we can actually you can actually share and use. Um and I'll speak a little bit about who we are, what we do, and how we did it, and what's next on the frontier. So, hope you like it.
[00:40] Yeah. So, who we are first of all? So, hi, I'm Marvin Namias. We're from Mexico. I'm the Chief Technology and Data Officer. Oh, and I'm Javier Hausel. I lead data and AI at Alpura. And where are we from?
[00:56] From Mexico. And who's going to win the World Cup? WE ARE. MEXICO. NO, WE'LL SPEAK A LITTLE BIT ABOUT ALPURA ACCOUNT. SO, let's move on. Yeah. So, a little bit about Alpura. Alpura is a major major powerhouse in dairy. So,
[01:13] we do yogurt, we do milk, we do protein shakes. It's been going on for 53 great years. One great thing about what we do is quality and we really love the the brand. Um we have many many products. So, you might say, "What do milkmen do here?" But, you'll see.
[01:28] Also, we we own our own cows. Yes. So, that's different from a bunch of CPGs that usually source that, right? From from external That is correct. So, it's a So, it's a cooperative at the end of the day. So, the owners are actually the farmers.
[01:43] So, a little bit I'll talk about our starting point. We're going to divide this in two parts. What we did in the first 120 days, believe it we can do it. There's partners here that can help you. Some of them are sitting down here. Um but, what we did in 120 days and then how we exploded doing apps the correct
[01:59] way. I know you've seen a lot of things this morning, but so our starting point we had over seven major platforms. And you see a little logo there. Well, Snowflake was one of them. Um so when we started this, we started this a little bit of about a year ago and it took us 120 days to say. But I'll
[02:15] tell you what we had. It probably sounds familiar. We had about seven and plus uh disconnected platforms. There were no data contracts. Uh that they were and different across BI tools. There were weekly reconciliations and 40 or more hours.
[02:31] You know what reconciliations are. I do a report. Sales sees it. We have to go over it. I don't agree. There was no golden source, basically. Yeah, so also what we used to have used to a few months ago uh we used to
[02:47] have very long development cycles. So once we would gather all the business requirements, uh all the user stories, it would take a really long time to develop the data behind any service, platform, or whatever.
[03:03] Uh and then the UI itself. Uh and after that validation. Um we also needed a bunch well, an army of people to actually keep keep this up and running. And uh our architecture, I must say, was all
[03:20] based on Power BI and Excel. So we would have people downloading stuff from the ERP system or whatever uh whatever they are using, right? Uh downloading that to Excel. Uh if Excel doesn't break down, then, you know, it's all joyful and and whatever, but
[03:37] usually what we do is operate on Excel and that entails having a lot of people uh from the engineering team spending time trying to make life easy for the Excel users. And then a a of very smart functional people
[03:52] spending time keeping Excel's basically alive, right? And I must say I am a very firm believer believer that uh Microsoft CEO Satya Nadella, he's CEO because of Excel.
[04:09] Not because of the hyperscalers. Because what we have seen is that there are a bunch of organizations still operating like this, right? So we we I mean we have not had the medallion architecture for maybe 7 8 9 years around.
[04:27] And still organizations struggle to have data that's ready to actually speed fast production cycles. Um Yeah. Keep going. Keep going.
[04:42] So what we did basically is tell the organization we're not going to continue operating in this fashion, right? Producing Power BI dashboards so that people can download then the data and then plug it into Excel and then VLOOKUP
[04:58] whatever they want to VLOOKUP and then produce another visualization or report or PowerPoint for whoever needs to make a decision. Uh what we did was basically spend those 120 days from the ground up transforming
[05:14] everything into bronze, silver, and gold. So preparing the entire platform so that we could bring about all the magic that we have seen today shown, right? In the keynotes. One thing one thing that was important
[05:30] is we we focused on doing the not sexy stuff. There's AI going on all along. Everybody wants to do agents. We said we're resetting the whole thing. We're going to do it in 120 days. That's 4 months, believe it or not. Uh so we basically moved all the data marts into
[05:45] Databricks and we selected the platform. So that sounds easy, but yes, you can do it, right? Well, what was the foundation at the end of the day? Obviously, the lakehouse. We put everything there. It was a unified platform, as you know. I'm not going to go over it. Medallion is very important. House is going to talk about
[06:02] it a little bit, but you know, people mention it, but they actually don't respect it, right? So, you end up seeing tables in silver that have gold and gold that's gold bronze and it's really not gold. So, we were really, really adamant and on trying to have these things because that's the source of truth. And
[06:17] the problem we had is you had sales fighting over with finance and fighting over the CEO. This number is not correct. It's not correct. Still goes on, but now there's sources of truth, right? Obviously, you need a catalog. I know it's not sexy, but when you govern data,
[06:33] then you're able to exponentially do stuff. So, we governed everything in Unity Catalog. Of course, we have an ERP. It's Oracle Fusion. We had migrated while we're doing these 120 days to Oracle Fusion, but at the end of the day, we were able to do data contracts and
[06:48] start putting data stewards and data data data people. Very, very good thing we're going to talk a little bit more is the semantic layer or whatever the marketing people call it these days, business semantics, semantics. I don't know the marketing people in Databricks, but at the end of the day, it's the formulas, right? We're going to talk about that because the
[07:04] formulas are embedded in reports. So, it's not enough just to move the data, but you need to move the formulas in a central place, whatever technology you want to use. And based off that, any agent or any person will actually see what every dot looks like. So, that's the semantic layer and what this really
[07:20] does, it sets you up for AI very, very fast. So, anything that Al Ali is saying this morning, you're able to turn it on and actually use it. If you didn't do this before, believe it or not, it's going to be a mess. So, that's our experience. Yeah, and I mean, I don't want to repeat
[07:38] over what's medallion. I think most of you know what it is. The main point here is emphasizing that uh we shouldn't be building data sets for every single visualization, for instance, or for every single process.
[07:54] The idea is to have a unified layer at the gold level uh where all the data can talk to itself, let's say, in some way, and then that something as powerful as as Genie, for instance, is able to access that data
[08:12] and construct or build whatever it needs to build without having us spend a bunch of time iterating with it. That's the whole point, but that is only achieved if you actually spend the time going through, polishing,
[08:29] or I would say refining data, right, from bronze to to to gold. And let me just be more pragmatic. Bronze, you just put the data however it is. We still get engineers saying, "Hey, let's just massage this a little bit before we put it in bronze." Just put in bronze whatever data you
[08:44] have from whatever system telemetry you have. We we we collect milk every day, so you can imagine the systems that we have, AS/400s and things like that. However it comes, just put it in bronze. That's the only thing. Silver, you take dedupe and you take nulls, and in gold
[08:59] you do aggregation. Okay, take that as a You know, you can have all these discussions, but people don't respect these three things. That's the most important thing in Medallion, just for the things you're going to end up doing. Yeah, the other thing that we did we have also spoken or we have not
[09:16] spoken, heard at the keynote, right, a bunch of stuff around context. Um to be able to steadfast whatever agents you have, there has to be context, and unless you want to have the same conversation over and over and over with
[09:32] your agent of choosing of your choosing, right? Um it's very important that part of the context is already up in Databricks. Uh governance is context. So, um I mean, I think also through the years
[09:48] we have uh been we have select, I think, uh thinking that governance is mostly access control. But then, doing the lineage, doing the descriptions, understanding where the data is, uh and you know, writing the
[10:05] context around that data, uh it's part of an enablement, uh precisely for Genie, right? So, Genie can, you know, you can produce anything you you want. It will go to Unity Catalog, take on the descriptions, and
[10:21] truly or fully understand what you're trying to say with a simple request. Um so, we had to we had to build this uh data governance office. We had nothing. Uh we had to bring in processes. Uh we had to reorganize ourselves also,
[10:39] so that people would become data owners and stewards, and really take control of of uh the data that well, basically, without rules, everyone's producing a bunch of crazy stuff that won't talk to each other, right? Yeah, we call him the sheriff, right? We
[10:54] have So, we have a a little pet called Manchas. It's a pet that's been out there 53. It's like Dr. Simi, if you've ever into Mexico, but it's a nice cow. It's a very nice cow. Um so, we call him the sheriff. But it very, very important, people don't want to be data owners, people don't want to be data
[11:09] stewards. But once you have the data correctly, I mean, we cleaned a lot of stuff. And as we were migrating to the RP, that really helped us a lot. So, this is not master data management. This is not master data. It's all types of data, operational and everything we talk about. So, having this in place makes
[11:25] you stand up for what it is. And nobody loves it, but you have to do it. Part of that as well was, yeah, the the the this the semantic layer, which we were going to talk about. Yeah, and the semantic layer again, it's a catalog just of a bunch of business rules, basic stuff. How do you calculate EBITDA
[11:43] or gross sales to net sales for instance, right? And it it looks like basic basic really basic stuff, but still businesses have a well, we have a ton of fights around how do we add and subtract stuff
[12:01] overall, right? So, if you want something as powerful as Genie, again, if you want to ask questions directly and that the entire executive team is seeing the same metrics and saying, yeah, I can trust that, you need
[12:17] to have a semantic layer. It's part It is part of the the context that you need, right? It's context engineering in a way. Um, and it's not only for Genie. I know we're mentioning Genie because it's very popular. We love Genie, but for anything you do, any report or anything do you
[12:32] that has to be in place. So, then suddenly after 120 days, uh Manchas One was born. We were able to migrate over 90% of our data assets. So, we have Nielsen, we have our ERP, we have our AS400, we have our SFA
[12:49] solutions, which are, you know, traditional channel mom and pops, how we take orders, how we're executing, but we also do the Walmarts, the Oxos, and the other stores that are in Mexico. That's very very important. So, all the data happens in Manchas One, right? Um, and it took us 120 days. We
[13:06] chose Databricks, and it's a unified data platform. So, after we had that, people were telling us, remember, 120 days, so what are you going to do next, right? Um, but you can do it in 100 days. That's the first part of this thing, uh but that's why we named it Manchas One. And everybody's going into Manchas One. It
[13:21] also serves as a integration layer, believe it or not. Uh, but we'll talk a little bit about that. So, that's part one, right? That's the platform. Yes, you can do it if we are milk people and if you're in technology company or anything, then you can do it as well. We did it with very few consultants, very
[13:36] very smart consultants, very very few people. You can do it yourself. It's just a matter of actually getting it done. Okay? So, we'll go to second part. Important part. The factory. I'm sorry.
[13:54] So, So, what the factory? Oh, go go go. Yeah. So, yeah, we had to build a factory. So, once you have all this refined data in place, the next question is what are we going to do with it, right? So, how do you get that to to users? Through apps or
[14:12] whatever else. Uh one thing that well, it freed our time to actually build stuff, right? Uh the first thing that we tried, well, it was basically by coding, right? Uh so, with the tools that we have now,
[14:27] everyone would be creating all sorts of uh code without a standard and different patterns and different ways of doing stuff. And of course, uh well, managing that was also very complex,
[14:43] right? Any bug, anything that uh you need to repair becomes complicated to manage. So, we started to bring in basically structure, right? So, basic guardrails, some standards, so that uh our
[14:58] developers, our engineers would at least attempt to try to do the same thing and have a particular order. Uh but then finally, what we did was actually bring in our own set of skills, guardrails, basically governance,
[15:16] so that anything that is produced uh through agents, for instance, or assisted by by agents, looks and feels the same without breaking innovation. Right? And that's that's our factory. Uh and we'll deep dive a little bit
[15:32] more. So, what the factory? What the factory is? The factory is the system around the AI, okay? And as House was mentioning it, we can vibe code, but if you've ever coded and somebody does some vibe coding at
[15:47] screen, you try to debug it and you're like, where's the log file? Why is the integration layer on the on the stuff? So, after we went to structure, like you very well said, um we were able to do patterns, based off this morning, Genie 8 minor but they
[16:02] they call it now Genie agent coder. Um they take certain specs, right? The AI dev kit, the AI dev kit. That's what we started doing in the structure, but in the factory and you're going to have access to this factory. It's open source. We have open source. It's hopefully it'll help you
[16:17] guys. Um we were able to put in our rules, our quota rules for anything that gets generated inside Databricks. It has our standard, the way it looks, the way it feels, and the way it does. So, this is basically is how how this is done is basically skills, our MD files. If
[16:34] you're familiar with Clojure or anything else, uh those we set up as a constitution. So, these are the unbreakables. If you're going to build something, you're going to build it in Python or you're going to build it in Node. Don't try to do it in R, right? With our logos. With our own logos, right? In the colors that we want.
[16:50] Don't you know, you guys are not UX, you know. You're great at doing reports, but don't do this. The same for interfaces. If you're going to interface to Databricks, if you're going to interface to our systems in the back that has an MCP server, you can do those specs here. So, if anybody's trying to use Genie code or cursor or tray or whatever they
[17:07] want to use, they're going against those rules and then we're generating those rules back. So, we're really doing a factory of things we can actually debug. And I mean, what what's the impact? What's the outcome, right?
[17:22] We can now produce I would say prototypes in a matter of hours. Hours. Stuff that would take weeks or maybe a few sprints. Now it takes just a few hours to and a few
[17:37] iterations. And I'm talking about you know more sophisticated stuff, right? Apps or complicated dashboards or that sort of thing. Um
[17:52] I have I have tried or Yeah, so I have tried to measure what would it take a really good engineer to to produce what we now produce with agents. It's around 15 to 20x in in productivity acceleration. So rather than a month,
[18:09] it's a day, right? Just a single day. The only thing is the new bottleneck stops being this technical part, right? So coding is not building the notebooks is not now a bottleneck. Validating whatever you're doing is not the bottleneck. As
[18:26] long as you know, you keep iterating and then returning from your coffee and and then keep iterating and iterating. you think is the is the thing that takes us the most time now? User requirements. Business rules. Yes. Business rules.
[18:41] Documentation. Customer that does net sales. Yeah. How do we calculate that? I'm just nodding to the finance and the sales people that are here from my report. Um but this is great because this is really puts down to your we're not holding you back. We were holding you back before.
[18:57] If you do the requirements, we're able to take you to speed. And that's the hardest part. But now, you know, we're able to do it basically, right? With really good documentation, a really good well-done user story for instance, with all the business rules, with the outcome,
[19:13] with the sources. We can now list the sources, right? Uh you can upload that to your favorite agent and it does, I don't know, in the first try at least 80% of the work. It's really impressive. Um and then yes, you have to iterate and
[19:30] And do do manual stuff a little bit. Uh not so much, but you you still need to do some coding, decoding, and that sort of thing, right? And and you would say we're the marketing people now. We're we're actually the technology people. We're going to show you some demo apps. We have over 90 apps already running. Uh we
[19:46] brought two. There's too many to talk about. Uh but I think we we've got two so you'll get a sense of of what we've done. I'm going to try to jump and do it live so you don't think we're doing it um uh and sorry about that. I don't have the cool gear that Databricks does in
[20:02] their demos. They type so fast. I they probably have a robot there. So the first one is I'm sorry for the Spanish, but we're from Mexico. So learn a little bit of Spanish. No, ambassador app. Talk a little bit about the ambassador app. So yeah, the ambassador ambassador embajador app
[20:19] easier. Um basically what it does is um It store by store it is able to tell you if uh what we're doing commercially is up to the picture of success of
[20:35] what whatever our execution should look look like in a particular store for a particular category uh if it's being executed correctly, right? But but I'll give you an example. So sales, they do that, right? They do merchandising. They go into a store. They put the product here. They put the product there. We've got all these great
[20:52] image recognition. We'll talk about that. But what we call ambassador is you're an Alpura employee. You're going to go to Che draw we. You're going to go to Oxxo. Do we see the Greek yogurt there? No, I don't see it. Take a picture. Send it and we'll do some AI on it. So I'm going to do a live See, right now it's
[21:08] detecting the the map. I'll try to make it bigger. Sorry. Sorry about that. So, it's yeah, we're in Moscone Center. There you know, I can move this wherever I am. So, I don't have to write it down. See, I'm not I'm going to I'm not going to write a store like you overwrite a store and say Walmart or Walgreens. It
[21:24] would take the rules on the back from that store. If I don't put anything, it thinks it's a detail channel or traditional channel which are mom and pop stores. So, I'm just going to leave George Moscone. I'm going to set up a file. I'm going to upload a file. The people that are here from Alpura have
[21:40] seen this photograph. From what store is it? No. This is a photograph that our CEO took. They're laughing because she always sends us this, right? This is a photograph of of something and basically I can just write the report and say there's no yogurt, no?
[21:55] But what's what's going on the back? If I click analyze with AI, it goes back, it hits an endpoint from Lama. It does the image and it goes It doesn't do image recognition against the picture of success. So, traditionally, you take that, you look at the picture of success, you compare the images.
[22:11] We're not doing that. We have the rules of the store and then the store It's taking a little while because the the whatchamacallit? The the spark Okay, came back. Now, they have micro sensors. out of tokens. Yeah, we've ran out of tokens, right? So, basically it's telling me for that
[22:27] store which is detail channel what at least we should be because I might say there has to be Greek yogurt here, but we don't sell Greek yogurt in that store or in that traditional channel. So, I can get an idea You know, yep, yep, it's cool, it's cool. Oh, by the way, I can't see the price. Those prices look very
[22:43] small. So, Thomas from RGM is saying, "That's great." So, whatever he enforces from revenue growth management, it can go. But this is the ambassador. We can take this obviously to the to the workforce which we have. And that's basically the ambassador the ambassador app. We'll talk a little
[22:58] bit about traceability. Anybody know what traceability is? I know it says lot of things, right? Especially if you're in data. For a CPG, the hardest thing is traceability. We collect milk every day. Okay, I used to work in Coca-Cola, right? It's not
[23:14] like, okay, people need more Coke, get more sugar, get more secret syrup, we'll do it. Cows give milk every day. So, how we do demand planning and what we convert it is very important. Then if you want buy one of our products, we can now today tell
[23:29] all the most to the cow level from where this milk came from. That's traceability for us. So, if we're able to change the narrative on this big stores, I'm saying big stores because things are happening in Europe and everything else, and you're able to scan your code, and you're able to see where that product
[23:45] came, locally farmed, from what cow, from what area. This is what we do. That's correct. So, so what we're doing is we're putting this out there. You you you can scan a photograph or enter what you see, you know, when you buy, you get the lot number, and the uh date. Let me
[24:02] just upload a photo. Um these are new protein liter milk. It has 54 g of protein. It's all natural. One ingredient. We don't add anything. It's going to come out with 60 g. I'm already giving you the commercial. So,
[24:18] as you can see, this is the photograph. So, we'll grab it. I'll just verify a product. Why it takes a while because if you know, it takes a little while to upload, but now that they have the micro apps um the micro it'll go up. So, what it's going to do is it's going
[24:33] to grab that lot, and it's going to do the traceability. But you would say, "Well, but you have an ERP, you can do that." I'm sorry, with all due respect. Anybody from ERP here? We love Oracle. We love SAP. They charge you millions of dollars to do a traceability app because you're connecting Tetra Pak
[24:50] with ERP, ERP with whatever you're doing, the ranch system, not the ranch system, it connects everything. What we did is because we had, remember, all the data where we have it, we were able to with Genie in a couple of of of
[25:05] cycles get all the information there and try to do traceability. And this is what we found out. This is actually Promilk. It is caducidad, which is it's going to be on the 0907. It comes from our planta in Cuautitlán Izcalli. You guys haven't seen this, but
[25:21] you're going to love it. And and what ranch it comes from. It comes from 57 ranches. We didn't want to get down to the cow, you know, alley cow. We don't want to say that yet. But for the consumer, this is very very appealing. So, we're changing their narrative. How we were able to do this basically is we're using models in the back that are
[25:37] doing the identification of of the image very fast. We have a couple of rules on that image recognition. And then we're hitting an agent that gets over 12 data sources from different systems in order to tell us from what ranch it comes. We can go to the level of the cow, but we don't want to do that yet.
[25:53] And it tells where it is and we have all the certifications. So, we're changing the narrative. So, what you're seeing here in a in a Databricks app is you have a fast API on the back because we put this on the public site. No, it's not an app. This is for a public site. So, it hits the API, the fast APIs.
[26:09] That's hitting an endpoint in Databricks. And that endpoint in Databricks basically is is doing all the image recognition, the agent and stuff. And by the way, as we're reporting this, I'm grabbing the geolocation from where you are. So, forbid, if there's a recall, I can know from what plant, from
[26:25] what cow. And if I need to recall that product, I can do it. But we don't use it for recall. We use it for traceability. So, people are really excited about this. We haven't launched it yet, but it's live. Also, I mean, the app, of course, as you search comes in, it's perfectible,
[26:41] right? But then, what's the most impressive piece is that this was built uh I don't know, 1 2 days for the first prototype. Um so it's very important to understand
[26:58] the tools that you have seen today uh in the keynote in particular are not for just, you know, an agent uh creating SQL for you. That's a really basic uh
[27:13] use case. You can start to create uh production-grade applications uh that actually serve a purpose for the business. It's not only insights, right? It's It's also It's asset building uh mostly.
[27:28] Yeah. So, yeah, the recipe for this First, get the platform right. Then, establish uh a governance office, a data governance office. Third piece, the semantic layer.
[27:44] Get the business rules right. Otherwise, it's impossible to scale anything. Um Fourth, guardrails, be AI-ready, and then when you have the factory, right? The What the factory? I'm What the
[28:00] factory? So, uh stop building apps one by one, right? Build a factory. So, yeah. Basically, what we did on the playbook. You can talk a little bit about this. And what you can do is, I mean, easy diagnostic, right? Where are you at? Uh
[28:15] what systems do you have? Where is your legacy? Uh How are you in processes, business rules, that sort of thing? How much How intense is your usage of uh Excel, Power BI, or whatever else you're using, G sheets
[28:31] maybe? Um And then, once you establish, well, what's your ground, uh what do you have to change, and that sort of thing, you pick your lakehouse. We do recommend uh Databricks, of course. Then, the next thing is get to build uh
[28:50] the medallion, right? So, try to establish as many the critical domains as possible as fast as possible. Um that's the first task of, you know, once once you get the lake house and your contract ready. Um then move on to the semantic layer,
[29:07] get all the business functions involved, start defining that. Uh and then finally, build the factory. Build the factory. Now, why we put this here? No, it's not a recipe. Obviousness causes blindness. So,
[29:23] rinse, repeat. It's very easy to jump from one step to the last step. Once we build the first part, we're exponentially building apps, we're exponentially doing reports, we're exponentially doing agents and multi-agents. So, this is this is our
[29:38] story. Now, this is the good part. We're giving you a way. It's for free. We love open source as well. So, we created the AI Genie factory. The GitHub is there. What you'll get is you'll get a shell and you'll get a couple of MD files. Seven skills, to be honest. The seven
[29:53] skills are I think you can see it up there a little bit on, you know, on the on the screen on the right. Uh and those seven skills you download, they're MD files, and those MD files uh you're able to say, "Hey, I am red, so I want everything red, and this is my logo." But, you can also say, "Hey, I like to
[30:09] do it in node. I like to do it in Python. Don't use Streamlit. It's nice, but it's not a HTML What do you want to do with HTML specs or instead of MDs? Um but also how you connect to certain systems is what you got to do. And this is what you can't do, and this is what you can do. You can't touch silver
[30:24] tables. You can only touch gold tables. So, when somebody opens Genie, uh it happens. But, what's the bonus here? There's a shell file there that call it's called deploy. You'll see the instructions. We wrote the instructions. If you hit deploy, you deploy these rules to your workplace or to your user.
[30:41] So, that means that anybody does it, they're able to do it. So, what was funny this morning, I was talking to Javier because it seems that every AI company that comes out, Databricks every year comes out with something and it's like, "Oh, what's happening with these seven startups, right?" They're These guys are getting into more and more and more.
[30:57] This actually builds upon what they were showing this morning. You can use the skill files in the agent file and apart from the rules from Databricks, these are the rules that you guys can govern apart from that. So, you're able to then suddenly be able to troubleshoot, debug,
[31:13] grow, scale, have security, do logging on the apps, on the dashboards, on the pipelines, and everything we do. We have a control plane on the pipelines because we register our logins differently. So, we have that in those rules. So, if somebody's going to build a pipeline, yeah, I'm going to do it, cool. I'm
[31:29] going to vibe code it and it's going to give me the best things from Databricks, but it's going to have my rules there. You can grab a copy of that and do whatever you want with it, but I'll help you. I I want to just be honest. And that's what the AI factory is. And well, I'll open it up for questions.
[31:45] Well, thank you. Thank you.

Learn more about the Databricks Data and AI platform.

The information provided herein is for general informational purposes only and may not reflect the most current product capabilities or configurations.