Consumer Industries: The Last Mile of Intelligence— Building Apps That Actually Reach the Frontline

Summary

Frontline workers in consumer industries—store associates, restaurant managers, and operations teams—make thousands of decisions daily, but the apps they rely on are slow to build, disconnected from analytics, and starved of real-time context because of the wall between analytical and operational systems.
Lakebase, Databricks' serverless Postgres, tears down this wall by eliminating the ETL pipelines, infrastructure overhead, and DBA bottleneck that have kept intelligent apps out of frontline hands, enabling real-time pipelines and gold-layer data to power production applications directly.
This video includes a live on-stage app build using a coding agent with seven parallel agents running simultaneously, deploying a store management app powered by Genie MCP for semantic layer consistency in approximately 10 minutes end to end—with RaceTrac sharing a real-world implementation story.

Consumer Industries: The Last Mile of Intelligence— Building Apps That Actually Reach the Frontline

Building an app has never been easier. Deploying an enterprise app— governed, scalable, connected to real-time data— has never been harder. Store associates, restaurant managers, and operations teams make thousands of decisions a day— but the apps they rely on are slow to build, disconnected from analytics, and starved of real-time context. The reason? A wall between analytical and operational systems that forces teams to maintain fragile ETL pipelines, over-provision databases, and wait months for production-ready tools. Lakebase tears that wall down. Databricks' serverless Postgres eliminates the infrastructure overhead, the ETL pipelines, and the DBA bottleneck that have kept intelligent apps out of frontline hands. In this video, we'll show what it means to build AI-first applications for consumer industries, walking through the end-to-end process: from use case design to live deployment, including a live app build on stage. Hear from brands already shipping Lakebase-powered apps, and leave with a repeatable framework for turning lakehouse intelligence into the tools your frontline teams actually use.
Talk By: Francisco Macedo, Execute Director of Enterprise Data, RaceTrac Inc ; Rob Saker, Global VP Consumer Industries, Databricks ; Roberto Robles Nacif, Global GTM CPG & Retail, Databricks ; Sathish Gangichetty, Lead Solutions Architect, Databricks ;

Chapters

FAQs

What is the 'last mile of intelligence' problem for consumer industries?

The last mile of intelligence is the gap between governed lakehouse data and the frontline applications that store associates, restaurant managers, and operations teams actually use to make decisions. This video identifies fragmented data, missing semantic layers, and lack of business context as the three biggest reasons vibe-coded apps fail to scale across large organizations.

What is Lakebase and why does it matter for frontline apps?

Lakebase is Databricks' serverless Postgres offering that eliminates the need for separate operational databases, custom ETL pipelines, and DBA infrastructure to power transactional applications. This video explains that Lakebase tears down the wall between analytical and operational systems, allowing frontline apps to read and write against live lakehouse intelligence without data synchronization overhead.

How does Genie MCP provide a semantic layer for consumer apps?

Genie MCP (Model Context Protocol) exposes a managed ontology that ensures KPI definitions and business logic are consistent across every application built on Databricks. This video demonstrates how Genie MCP acts as a unified semantic layer so that store management apps and loyalty applications return the same metric definitions regardless of which team built them.

What did RaceTrac build using Databricks?

RaceTrac built a store intelligence solution on the Databricks Data and AI platform that surfaces real-time operational data to store managers through a governed application. This video presents RaceTrac's implementation as a reference architecture showing how a real-time data pipeline feeding a gold-layer data model can power a production frontline app built with the help of Genie Code.

Full transcript

[00:07] Thank you everyone for being here. I'm very excited to talk to you about building apps that actually deliver and get to the front line. Next one. My name is Roberto Robles. I'm the global go-to-market lead for consumer goods and I'm going to be
[00:19] sharing the stage with Satish. >> Hey guys, I'm the elite solutions architect in retail CPG and travel. I primarily focus on AIML and so today you'll see a lot of the
[00:31] coding agent stuff around building apps on Databricks. >> And with Francisco and the racetrack team. So we have a pretty packed agenda. We're going to share a couple of slides. We're going to do a couple of demos. Then we're going to have the racetrack
[00:43] team come to stage and actually share a real-life example of a an app that they built using Databricks. And then towards the end there's going to be a QR code, very important QR code. If you scan that
[00:55] QR code, you're going to be able to schedule an actual workshop, in-person workshop with Databricks team. So if you're interested on executing or delivering, you know, what you are going to see on the screen, just scan that and let's make sure that you schedule a
[01:07] meeting. All right, so today's super easy to build apps. There's many, many, many applications out there that allow, you know, business users and technical users
[01:20] to build data applications using natural language. You can use plain English to create pretty cool and amazing looking apps. However, the reality is that it is extremely hard
[01:34] to scale these vibe coding apps across organizations. And the three biggest challenges that we see, you know, across the industry are, first and most, if your data is still fragmented, then a
[01:47] lot of these applications still have a lot of incomplete data. The second biggest challenge that we see is the lack of semantics. So as you as your organization is building different data apps, if you don't have a unified semantic layer, then it's very hard to
[02:00] have that unified single source of truth where the KPIs mean the same no matter which application is being used. And the last one is, you know, it's very hard to infuse that business context on how your
[02:13] business is run on these five coding applications. So, what is required to build trusted AI? First and most, you have to have data management. You You need to be able to
[02:25] unify You need to be able to unify all your structured and unstructured data into a single in a single place. Then you need pretty good data governance. So, you should be able to control who can build data apps, but you
[02:39] should also be able to control who has access to what. Then you need a AI building layer, meaning you should be able to infuse GenAI and agentic into these data applications.
[02:52] You need fast engines, and the meaning of these is you need to start thinking of moving beyond just using apps to surface insights into how can you use apps to actually drive actionability across your business.
[03:05] And lastly, you also need a pretty good user experience. At the end of the day, you want to build an application that is used by your business, and you want to drive adoption. So, this is where our platform is perfectly positioned to help you go
[03:19] end-to-end and build applications that can truly help you scale and deliver value across your organization. I mean, all the different models, I think you heard a lot about these during the keynote, but we're perfectly positioned
[03:32] to help you go end-to-end from the data all the way to the last mile to surface these data apps to to your users. Next one. So, the example that we want to use for this session is called the
[03:44] Monday morning report in retail and consumer goods. So, if you're in retail and consumer goods, uh And probably know that Monday morning is one of the most important days of the week because that's the day of the week where, you
[03:56] know, you plan your merchandising strategy, your assortment strategy, you know, your category strategy. So, um there is a lot that needs to happen and unfortunately a lot of organizations are still using silo uh silo BI and
[04:10] Excels to run this important report. So, we're going to actually showcase what good looks like. So, this is a full end-to-end app built within ThoughtSpot.
[04:24] And the idea here is imagine moving away from a static power BI or a BI report or Excel into a dynamic application. And what I want you want you to pay attention to is some of the
[04:36] capabilities, right? So, first and foremost, you're combining uh visuals, you know, users are still like to use see visuals, but then you also have the Genie capabilities. If you see here, I see that my total sales are down 11.2%
[04:52] versus my plan. I can just come here and ask Genie, Whoop. I don't know what's going on here. Of course, it's a live demo.
[05:04] All right, so I'm going to ask another one. But the idea here is that you can have pre-populated questions based on metrics and KPIs that you have in your homepage. So, the user is not just asking plain questions, right? You can guide the user on which questions they
[05:16] can ask. I'd be worried without because the internet is a little slow right here.
[05:33] All right, I'm going to just uh keep going, but you get the the the the picture is you have visuals combined with Genie, so you guide the user into which questions they can ask and which questions they they should be asking. The other piece that we're seeing is specifically in retail and consumer
[05:45] goods is something that is very hard is to bring together all these different signals into a single place. So, if you think about Monday morning reporting, you have inventory, you have e-comm, you have reviews, and you have sales. So,
[05:58] it's very hard to bring all these different signals into a single unified, you know, application or visual that your users can consume. So, here you can see that you have very important KPIs. You're combining all that. Uh you're
[06:11] highlighting top movers, you know, where the users need to pay attention to. And lastly, you know, you have your as genie. So, you have a full, you know, bake-in talk to your data interface uh built
[06:24] within the application. So, now what we're going to do is uh Satish is going to show you how you can build once you have the foundation that I talked about. Once you have the data, the governance, the semantics, he's going to show you how quickly you can actually build and
[06:38] replicate uh something like this. So, Satish, I'm going to pass it off to you. >> Awesome. Thank you, Roberto. Uh show of hands if you already used coding agents of any kind, you know, Claude, Codex. Amazing. This is what we want.
[06:50] Um show of hands if you used Databricks apps with the coding agent. Show of hands if you used OmniGen. Well, one person. Okay. So, here's what we're going to do. We're going to try
[07:02] and hit a stretch goal today. We're going to go slideshow. We're going to try and replicate that Monday morning app that Roberto showed. The whole thing was based on Oh, you know what? When Anthropic released this
[07:14] uh this model, which is now non-existent. Uh unfortunately, um they made Fable write its own launch video, right? I was like, "Okay, so maybe we can take inspiration from this and do basically something similar for
[07:28] the session." And but then Fable passed. And now we're back to Opus 48, right? So, we're going to build this entire thing on Databricks. The most important thing here is that for you to do anything like
[07:40] this, you need to set up a harness. What that means is that your coding agent, your Codex or a cloud code or whatnot, needs to have access to things that a normal developer would have access to. Like Databricks CLI uh to be
[07:53] able to commission resources on your Databricks workspace. Uh the model itself coming from the uh Unity AI Gateway. Skills about Databricks agent skills. Databricks now publishes this thing called Databricks agent skills. It's a
[08:05] publicly, you know, surfaced GitHub repo that you can attach to practically any project that you're building on any of these coding agents. The deploy target, of course, is Databricks apps. And then we have app state, meaning the the
[08:17] short-term memory, long-term memory stuff being written to lakehouse. And then for the MCPs itself, this is where the secret sauce happens, right? You heard Ali talk about uh Genie ontology today. We're going to use the Genie MCP server to surface up that
[08:30] ontology on demand. And then we have other MCPs to look up, you know, code bases, query them, uh do some research for us. That said, what I want to do is I want to set this off in the background. So, what you see here is Omnigen. And
[08:43] I've been timing this whole thing. Uh firstly, what I'm going to do is I'm going to open up a new shell. And I'm going to just post.
[08:56] Oh, one sec. I'll just post a command on it. And then say go. Uh this is going to start off a new cloud session. What it's doing is it's called loop engineering. Show of hands if you know loop engineering. This is the brand new thing
[09:09] where the idea is to not prompt the model, but you give the model uh completion state, meaning a success state, a success criteria, and the model, in this case in the agent, is going to go build out, you know, over
[09:22] time, dispose of maybe agent teams or maybe dynamic workflows, and decide the role and the goal and the backstory for this agent and build these things out in parallel, right? So, I'm
[09:34] going to I'm going want and do this very quickly uh and launch this workflow and head back to the slide deck to see how we actually got to this point, right? Um Okay. So, run the workflow.
[09:48] And it's going to launch workflows in the background. You see there are like seven agents in the in parallel that's running. So, let's head back here. So, this is the coding agent harness. We've fired off the coding agent and it's doing its work in the background inside
[10:01] of Omnigent, which is like using cloud code behind the scene. And this is what we're going to use to build a Monday morning app. The most important thing in this loop engineering process is that you need to set chase goals. And if you need to
[10:13] chase goals, the first thing is you need to have clarity around what you're chasing. Just like this this sounds like the thing. Good part is Roberto already has the finished product. I can just do the same thing that the guy who, you know, presented Fable 5, like made his own
[10:26] video thing. I got a video of like what he presented. Uh and extracted out the requirements, and this is what you see out here, all the requirements by uh the product requirements, the dashboard requirements, Genie
[10:39] requirements, and all of that. Again, there's no human in the mix here. The The agent is doing this by itself. The second thing that I want to make sure of is I want to deploy this to a database you know, app, which means the coding
[10:51] agent needs to know which workspace, what resources should I commission, uh which MCPs do I have to use, what is this lake base instance that I have to spawn. All these things are super important, right? Again, don't overthink
[11:04] it here. Absolutely important to make things simple. If it's confusing to you, it is going to be confusing to that coding agent. Right? Try and keep this as simple as possible. Then, we bake the environment. You know, of course, when we need to, you know, usually in your
[11:17] situation, you're going to have all the data available to you. But in in here, I'm having to commission the data, synthetic data, set it up so that I can surface this in a completely different workspace. Uh but
[11:29] with all that being said, this goes into defining that spec doc. The spec doc is the termination success criteria that the goal has to hit. So, as long as the success criteria is met, uh this this coding agent processes, you
[11:42] know, is going to go to run uh an iteration. And so, that's what we ended up with. So, we had this goal, uh which is baked into, let's say you're using Codex or Claude code, both of these have this. Uh and essentially, we trying to deploy
[11:55] this Monday morning uh retail insights app, and it's built for that spec doc. And we're going to trigger dynamic workflows. Was like Claude is going to figure out, okay, what do I do here? And oh, yeah, so hold on 1 sec.
[12:08] Just going to let it um define the workflows. And um it's running in the background. You can see it's doing seven parallel agents, and it's figuring out what needs to be what needs to be done here, right?
[12:20] And then I set up this like last word here, ultra code, which is to make sure that Claude uses the correct amount of reasoning to ensure that it's defining the workflows properly. So, when we do all of this, what ends up happening is
[12:32] that you commission these agents in parallel. This is what when people talk about autonomous agents running things on a loop over time against a goal, this is what they're talking about. And it just runs until the job gets
[12:44] done. Why do you want to do this in a goal? Because there are two parts of a goal of what you need to accomplish. One is around the generation of the code, another one is adversarially saying, "Hey, this is not
[12:56] done. This is not good enough." Right? So, then there's this whole back and forth between the generator and the evaluator. For the data science people in the room, this concept dates back as like GANs, which are now like almost
[13:10] Jurassic era ideas, right? Um So, and then over time, it just runs this thing, and eventually, hold on 1 sec.
[13:23] Yeah. Eventually, it builds out the whole thing automatically, right? So, uh so far, we've spent 4 minutes in the goal and it's going to keep continuing. And then when you finish the goal itself, here's like a one-shot result of
[13:35] what what So, you before I got to this demo, I was able to run this whole thing multiple times over. It was anywhere between 10 minutes to 20 minutes. Of course, I just spoke for like what, 5 minutes?
[13:48] And we're almost at the like finished stage here, but I want to try and take some time to walk you through some of these components that go into doing uh what exactly this is, right? So, I talked about, you know, spawning the goal, which we all saw up here.
[14:01] Uh but then we also have the spec doc. So, if you go to the spec doc, line item uh sub, you know, section six talks about what the end criteria needs to be.
[14:15] This is what it's trying to measure itself against. And the goal's running right now in the background for 5 minutes and we should probably be finished around the 10-minute mark uh or the 12-minute mark. But I've already done this previously so
[14:29] that we don't have to be sitting around here waiting for this goal to finish. Um and so you can see the whole thing ran when I spawned the goal all the way down to the build time
[14:42] is about 11 minutes, right? So, it gives me this app. If I click through it, it takes me straight to the Monday morning app. Right? It does the exact same thing that Roberto's talking about. It uses the Genie MCP. How do I know this? Well, I
[14:55] triggered this out. And let's go to the workspace and see if the Genie MCP is actually resolving it, right? Um
[15:14] Okay, so go to Genie one. And you should see here based on the question that we asked, Genie is doing this analysis live without any other context, right? So, as
[15:27] one as this this guy finishes here, it should hydrate the upstream app. And this is how you can take advantage of the ontology. Bring every type every kind of document into Data Bricks, structured data, unstructured data, your
[15:39] functions, your tool calls, your MCPs, everything. And then that then builds out that managed ontology for you, which you can then directly access through that MCP. Right? So, when it's done, it's going to go show you how it did it, what the
[15:53] response was, what the results were. Also going to link to that same exact result that you saw on Data Bricks, right? So, very simply and also that not just that, we were able to build out the whole
[16:06] you know, the map view and everything else that Roberta was talking about um all by putting your coding agent, equipping it properly with the right skills, the right tools,
[16:18] uh ensuring the harness engineering happen properly, and then we set the loop in motion and the loop would go in and get the work done, right? So, if you if we did this right, it would say the goal is achieved, and this was the uh
[16:30] the other session, but in this other session, this still this thing is still running 7 minutes in, right? So, I don't want to make you guys wait, so what I'm going to do is I'm going to go back here.
[16:44] And I'm going to just quickly summarize what we saw, right? So, you set a goal in motion uh and by saying, "Hey, you want to deploy this to Data Bricks apps and give it all the skills, all the
[16:56] uh tools that you need." Uh and eventually, over time, it is going to go do practically all of that stuff, deploy that into Data Bricks apps. Uh this thing took about what, 9 minutes
[17:09] 41 seconds here uh in about 10 minutes. And then that's the damn demo that we actually saw. It obviously did not get everything right. Look at the database
[17:21] icon there. Like it's not right because it does not know what the icon is. But those are minimal things you can iterate on. It's not really that hard. Okay. So, >> Yeah, so just to recap, right? Once you have the foundation done, so a lot of
[17:34] our customers have already spent a lot of time building that foundation, getting the data right, getting the semantics right. Now you can see how you can build a full end data application in less than 11 minutes.
[17:46] So, this is where we're seeing that the next generation of analytics is no longer going to be silo, you know, BI. It's how do you take the most of GenAI, a GenAI and analytics, and how can you
[17:58] surface these to your business users? So, with that, you done? >> Yeah, it's it's the goal still running, but it will be done in like a couple minutes. So, I'm going to I'm going to have our guest Francisco come and come on up
[18:10] and talk about this. Maybe this can run in the background and we can come back and check, right? So, the deploy succeeded already, as you can see. It actually hit the goal without even me doing anything. Um but it's going to do the, you know,
[18:22] final work of like getting the feedback from the manager. And every everything goes well, we'll get like a URL, and if you click the URL, it takes you to the app.
[18:58] All right. Finish up. >> Hey guys, my name is Francisco Osito. I work for RaceTrac. For those of you who live in the South, it's a It's a pretty large um, convenience store, gas
[19:12] station. Uh, RaceTrac also owns Gulf Oil. And, uh, I had the the data infrastructure uh, division of RaceTrac. And, uh, I was invited by Databricks to come talk
[19:25] about the, uh, store intelligence, uh, solution that we built. And, uh, you know, I I I like to to sometimes make the analogy, for those of you who who scuba dive,
[19:37] this is truly like, you know, s- going from snorkeling to scuba diving with a tank. And, this is truly what we found. It is remarkable, the pace and how quickly we could port a
[19:50] Databricks um, a, uh, Power BI dashboard that just wasn't working into Databricks app. And, uh, and this is kind of how this this this whole thing started. We we needed
[20:03] to build a dashboard that was really, uh, uh, near real time with lots of data that would allow the the executives to ask questions to the
[20:15] data and view, uh, KPIs and metrics that we just we just couldn't deliver, uh, on our old platform. It we had all kinds of issues, the data finishing up, the data sets
[20:27] loading. It was just such a nightmare. And, uh, and then the the wonderful Databricks team helped us, uh, with the solution just about maybe two or three months ago, and it's live.
[20:39] Um, so, I'll you know, just through the the process of, you know, how we got there, right? Why we started, uh, what was under the hood, the architecture, and how AI accelerated our our solution.
[20:53] Um, and this is, you know, the problem. What was the problem we're trying to solve? Executives are waiting for data on the dashboard that that under our previous architecture and design just wasn't
[21:06] getting done. And was fairly stressful when every other day your your data is not loaded and and folks are waiting for it and they're just not there. Um so we um
[21:20] Yeah, we briefly talked about this, right? We we just couldn't the solution we had just just didn't didn't work. Um there were different data sources. We we had to find a solution that that sort of combines all our data points into one
[21:34] source of the truth. And uh and and bring it all into a centralized location was the solution. Um we we we reversed engineer this this Power BI dashboard using Gennycode.
[21:47] Um that was kind of the the the starting point um for for our our dashboard. Probably 2/3 of what we had was pretty easily ported over. We we we sort of
[22:00] fixed the the rest of it as we went along. Um these are just some some data points. We you know our old solution um had 18 pipelines.
[22:12] Um was Owen Bronze. Well, actually this was the the the the original version of the dashboard. We made improvements um throughout the process. But essentially we we had it was just a
[22:25] complicated solution. And we had to simplify it and and this was the the sort of the solution, right? We wanted to have um some some golden path as to how we built
[22:38] it. Um the we used the the the the Databricks app, Genny Spaces, and uh metric views. That's pretty much how the solution was built. And then we we denormalized all all the data that we
[22:52] needed for this dashboard into our gold layer. And uh and that that's what powered our our um uh dashboard which I'll I'll show you guys. So this was kind of the the process that we went through.
[23:05] Um metric reviews became our source of the truth. So, everything that that we you know, all our KPIs, all data, we we needed to be able to process that data very quickly in into a single normalized
[23:18] view. And you know, I guess you know, if you want to compare to how we had it before, this guy was kind of equivalent to having a you know, different a huge data set in part of BI that just never loaded. And
[23:32] in metric reviews just just was just the way to go. Um So, this is just a little summary from from bottleneck to govern on your real-time data, one certified
[23:44] source of truth. And and this was a really a a good opportunity for us to really go back and work with the business in terms of of you know, the the whole process of data ownership and and and
[23:57] and partner with them to actually uh clearly define some some data metrics that that weren't correctly defined in the past. But this this was one of those things that when we started developing the
[24:10] solution, um we thought we're like, you know, this is going to take probably three full resources for at least 2 months for us to pull all of this together. And this was us thinking how we built it before.
[24:23] And as we introduced Genie code into the mix, um it was like, you know, it's life-changing. We we didn't realize that the whole solution could have been done uh probably in a week with a a lot more reliable
[24:36] data. It was just you know, going back to my scuba dive uh analogy, it was truly that. Um let me let me bring up the app so you guys can see what it looks like.
[24:50] So, this is All right, man. What is your
[25:05] >> That didn't work. Yeah, so this is this is the actual this is the actual app. This is all test data. None of this is certainly true. So, your Mac is not
[25:17] touch. No, no, I don't You got it. Got it. Got it. It's all good. Um yep. So, the the the way the this is the uh we wanted to be able So, essentially the
[25:31] idea behind the solution was we wanted um you know, an exec from data from uh Racetrac to walk into a store and actually zoom into a given store
[25:43] and and go, you know, have metrics about that store. The the uh the next phase of this which we're in the process of porting over to to a phone app is we we're going to
[25:55] enable geolocation so that, you know, folks who own Racetrac they want to walk into a store and be able to pull up their phone. Geolocation will will then kick in and they will be able to see
[26:08] every single KPI from every single store that Racetrac owns in real time. And uh and this is what we're working on now. I think I even have a a picture of what what it's going to be beginning to
[26:20] look like on the on the app. And and obviously, you know, the the we wanted to be able to ask the the uh the data questions. We wanted to be able to allow executives to you know, ask
[26:33] specific questions about the data um on the on the dashboard. And now we're actually porting that over into the app as well. Um and and this was the goal and this is something that we could not have done 6 months ago.
[26:46] And truly, you walk into any store, any Racetrac store, and this will geo-locate. You can ask any questions that you have uh about your data, and this thing just works. So, this was a This was kind of the the final solution.
[27:07] >> Awesome. Thank you, Francisco. >> This is it, right? Yeah. >> All right. So, we uh 45 minutes, you know, we walk you through the foundations. We actually built an app, and I heard an actual customer that has implemented this in
[27:20] production. >> If you can go back to the >> I just want to say the app app build is done, so we'll go back to this. >> Yeah. >> So. Here.
[27:37] >> Uh so, this is the QR code that I was mentioning. I mean, we from a Databricks perspective, we would love to help you, you know, replicate and help you be successful on any use case business use case that you have. So, the call to
[27:49] action here is if you want to scan this QR code, we can come to you, and we can help you with a whole day workshop, and we can help you with end to end, right? Talk about the state of your data, you
[28:01] know, how's your governance, uh Francisco talked about the semantic layer. I cannot emphasize enough building the UI. You saw it's very easy. Building the user interface can take 11 minutes. Building that semantic layer,
[28:14] that's where the where the secret sauce is. But, we can come to you, and we can help you build that end to end all the way to the delivery aspect of it. So, if you want to scan this QR code, uh you're
[28:26] going to have a little form. Reach out to us. More than happy to to help you be successful, and you know, deliver value with the platform.

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.