Meet Bricksy - The Hyper-Personalized, Privacy-First AI Copilot for the Software-Defined Vehicle

Summary

Bricksy is a voice-enabled AI copilot for software-defined vehicles that streams live vehicle telemetry—including engine temperature and tire pressure—into Unity Catalog via Zerobus for real-time anomaly detection on the Databricks Data and AI platform.
The agent uses Vector Search over car manuals and service history for precise question-answering, and Lakebase to persistently store per-driver preferences like ambient light settings and driving mode across multiple users of the same vehicle.
This video demonstrates how manufacturers can deploy proprietary AI tools directly on the lakehouse, ensuring sensitive telemetry never leaves their governance boundary while also enabling post-hoc driver profiling for insurance risk assessment.

Meet Bricksy - The Hyper-Personalized, Privacy-First AI Copilot for the Software-Defined Vehicle

As automotive industry pivots to the Software-Defined Vehicle (SDV), the challenge has shifted from collecting data to interacting with it in real-time. Drivers demand a seamless, personalized experience, while OEMs demand data sovereignty and low-latency interventions.
Bricksy, a next-gen AI agent powered entirely by Databricks is an android app. We'll demonstrate a unified architecture for building voice-enabled driver companion that thinks, remembers, and protects having:

Real-Time Nervous System: Vehicle telemetry (Engine Temp, TPMS) streams directly into Unity Catalog via Zerobus, enabling near real-time anomaly detection.
Infinite Memory & Context: How Bricksy utilizes Vector Search and Lakebase to instantly recall car manuals, service history, and long-term driver preferences.
AI That Comes to Your Data: How manufacturers can deploy proprietary tools and functions directly on the Lakehouse, ensuring sensitive telemetry never leaves their governance boundary.

Talk By: Hussain Jagirdar, Solutions Architect, Databricks India Private Limited ;

Chapters

FAQs

What is a software-defined vehicle and why does it need an AI copilot?

A software-defined vehicle (SDV) is a car whose features and behavior are managed primarily through software rather than fixed hardware. As vehicles generate continuous streams of telemetry data, an AI copilot can interpret that data in real time—diagnosing issues like tire pressure drops or engine overheating—without requiring the driver to search online or visit a dealer.

What technology stack powers the Bricksy AI copilot on Databricks?

Bricksy is built on the Databricks Data and AI platform using Unity Catalog for governed data storage, Zerobus for real-time telemetry ingestion, Vector Search for car manual and service history retrieval, and Lakebase for persistent long-term memory of individual driver preferences. A supervisor agent coordinates these components through MCP tools.

How does Bricksy handle multiple drivers sharing the same vehicle?

Bricksy uses Lakebase as a persistent memory store to track individual driver preferences—such as preferred ambient lighting, seat height, and driving mode. Each time a recognized driver enters the vehicle, Bricksy recalls their preferences automatically without requiring manual reconfiguration, even when several household members share the car.

How does Bricksy protect sensitive vehicle data within manufacturer governance boundaries?

By deploying proprietary tools and functions directly on the lakehouse, manufacturers ensure that sensitive telemetry never leaves their governance boundary. Unity Catalog governs access at the data layer, while the AI agent operates entirely within the manufacturer's controlled environment rather than routing data through external consumer phone platforms.

Full transcript

[00:08] Thank you all of you for joining. Today, I would be presenting the Brick C. So, just to give an introduction, my name is Hussein and I'm a architect for Data Bricks. And so, we have been working with one of the vehicle manufacturer in India and we
[00:22] have been making some progress in making a co-pilot, actual co-pilot, which is a hyper-personalized privacy-first AI co-pilot for the SDVs, which is software-defined vehicles.
[00:35] So, uh quickly going through the problem statement, why was there was a need for a personalized co-pilot, right? So, if you think from a diverse perspective, right? So, driver want a quick quick
[00:49] command functionality. Also, basically, right now, whatever the bot that you have in the car, be it a Gemini or a uh phone phone bot, right? It's very much limited to the
[01:02] capabilities of your phone, right? Whenever you connect via Apple CarPlay or a Android Auto, right? Either what you can do is via you can take some maps or music, that's all, right? And it's more of a
[01:14] command-like interface, where you give a command and it gives a response back to you, right? So, we now the manufacturers are wanted given the AI that we have the AI uh frontal
[01:28] models that are working so good, the manufacturers also want to build a bot inside the vehicle, which can converse with the vehicle telemetry data. And when I say vehicle telemetry data, it could be like, if there is a problem,
[01:41] let's say let's say in the tire pressure of the vehicle, you can directly converse and check, "Hey, what is the tire pressure of my car? Is it going down? Whether the engine temperature is going up?" And all of this data related to the vehicle should
[01:54] be accessible to your bot, right? And you can directly converse with it. So, one example that I gave in the in the third point is the works with the car telemetry data. So, all the safety related questions you can directly
[02:07] converse with the with your bot rather than searching on the Google and all. The fourth point again becomes very important when you want your car or your bot to remember the long-term preferences, right? So, in case in
[02:21] typically household car is getting driven by a like two, three folks, and we want that during the conversation whenever you are setting some preferences, right? Either ambient light or driving style, seat height, all of
[02:35] this personalizations can be remembered by your bot so that each time you enter the vehicle, you don't need to set it manually. Or say that, "Hey, I prefer sports mode." Or I like the 18° C, something like of temperature, right?
[02:48] So, and also the fifth point becomes very important, right? So, whenever you want to do a uh uh natural conversation with a bot, it becomes very important you do a multi-turn conversation, right? So,
[03:01] like I gave you an example of a command like instructions. You don't just want to give a command that, "Hey, do this." And you take a response. Sometimes you want to do a more deeper conversation, maybe saying that, "Hey,
[03:14] from place A to place B, what is the best route right there, right?" So, it might give you a map location. And then meanwhile you also want to search, "Hey, do you think are there any theaters which might be showing some place, right? Or some movies in that particular
[03:28] route, right? So, in that case you should know what the previous conversation has been going to. What we typically do in a chat GPT as well, right? In a session. Now, if you think in the manufacturer's
[03:40] point of view, it becomes very tricky to how would you handle this whole telemetry data, and then you can use it to your AI system backed for all the analysis and the actions. Number one.
[03:52] Number two, also if you think about the case where is there is there any portal available right now which I can directly update all my car manuals, directly upload my car manuals, or maybe servicing logs, so that in case
[04:06] let's say after servicing if I get a if I'm getting some voice from my back tire, I'm just just can ask my car that hey, is it because of the suspension that got changed in my last servicing?
[04:19] Is it because of that, right? Otherwise, you have to you consult your vehicle expert or maybe call back and do the analysis all those sort of things, right? So, the third point becomes very
[04:31] important even if you do some sort of a work around to integrate, it becomes very important that the privacy and security for the manufacturers becomes very tricky because they don't want to send all of this vehicle telemetry data
[04:45] to the tech giants, right? Or the frontier models. owners, right? They want all of this AI to be embedded in their framework. And that's what the fourth fourth point is that hey, we need
[04:58] a unified framework where we can bridge the structured, unstructured, and also the real-time telemetry data, right?
[05:10] So, for all for this problem statement, we have kind of introduced a Brixy. So, this is the architecture diagram of the what the Brixy looks like, right? So, in the left-hand side, you see there are three drivers which are kind of
[05:22] interacting with the same platform or a same uh user interface which is your voice copilot, right? Similarly, when you you might be entering the car or multiple drivers might be using your same car,
[05:35] they might be interacting to the same copilot and which is either a Android client or maybe a Apple whatever it is, iOS, and it basically the task of it is to understand what the what's the
[05:47] command is, what's the instruction is, and do the speech-to-text and text-to-speech back back to you. And all all the green things that you are seeing are the sensors, right? For example, the tire pressure monitoring
[06:00] tire pressure data and the engine temperature data. All of this data are uh uh should be flown in the real time uh real time uh streaming in a streaming fashion to a a
[06:13] lake house or a lake data warehouse that we have shown in the right-hand side if you see is the Unity Catalog, right? So, all the sensor data all the telemetry data is kind of ingesting into the Unity Catalog following the medallion
[06:25] architecture of the bronze, silver, and gold via zero bus ingest, right? Uh in the middle of this thing you this is the Bixby agent which is sitting right in between which is interacting
[06:37] with your Unity Catalog where all the telemetry data is coming in. And in the right-hand side if you see is it's the lake base. Now, in the lake base what what's we are using for the lake base is for storing the preferences. So, if you can see each
[06:52] of the driver Ali, Bella, and Chloe has their own sessions which are maintained in the lake base separately where we are storing the long-term preferences of Ali, Bella, and Chloe, right? And also
[07:04] some short-term history so that each session can understand what exactly the what exactly the conversation is going on, right? In the top most it becomes very important that the vector search is
[07:18] kind of embedded directly into the Bixby. And why the vector search is? Because since we have the unstructured documents of the car manuals and and your servicing logs, right? So, is there any a framework where
[07:31] while buying your car the manufacturer can ingest that for particular model your particular car variant this is the car manual so that it your the agent knows what exactly the ideal
[07:44] temperature should be looks like or the ideal this all the parameters looks like and after each servicing maybe there is a portal where there you can directly ingest all the servicing logs back into
[07:57] the vector search so that after a day's time when you're interacting with the AI bot it has the all the relevant information and it preserve the context. Right? And again if you see over here this is
[08:10] the health diagnosis tools why we have captured the health diagnosis tools is all the different manufacturers has their own protocols and set of actions they want to take in
[08:22] before jumping into the conclusion, right? So these are some deterministic set of rules and functions that can be embedded let's say in the unity unity function tools that can again attach to the brick
[08:34] seat. So your brick seat is more of a supervisor agent. Right? And it's my it's interacting either to the lake base for the memory for unity catalog for the structure data for the unstructured data it's interacting with your vector search
[08:48] and obviously it might be also cooling calling some tools in between for the analysis. Right? If we go to the right hand side like it's more mostly about what is the action that is getting taken, right? So
[09:01] for the actions what we are we I'm kind of we I would be showing the demo as well. A simulated dashboard obviously we can't show you a car here, right? But MCP tool
[09:13] how we can connect with to the simulated car dashboard so that if you can see that hey I'm feeling a little warmer, right? So when you say I'm feeling a little warmer it should know that hey right now what is the temperature right now or
[09:25] set maybe it's set to 20 degree maybe I need to make it down to 18 degree based on this preferences, right? So the the output should be shown the some action should be taken now, right? So second part is the dealer management system.
[09:39] Sometimes there is already a good dealer management systems already available in the market which are used by the while calling you can book the servicing slots, right? But if that APIs are
[09:52] available, can we embed embed them can we wrap them with the MCP servers and can we can connect to the brick agent. So, this is the ask is like whether can I use my brick see to
[10:04] direct to converse with it and say that here I'm feeling that something is not wrong can we book a service appointment. So, directly it will check that here in which region do you want to check and these are the servicing these are the
[10:17] slots which are available for this particular date. Do you want me to proceed? If you say yes, then it will book us slot for you, right? And of course the last part is more of a post talk analysis where since you get all
[10:30] this telemetric data, it becomes very important that you can take all of this data to do the user profiling. And when I say user profiling, it's more of a driver profiling where for each
[10:42] not exactly driver driver but for a vehicle profiling that here this vehicle what is the average kilometers getting drawn and what is the variation and that all this analysis of is it how good this vehicle is getting drawn
[10:55] can be used either by the other agency for selling the car or maybe for the insurance as well to calculate the risk premium, right? So, yeah,
[11:07] these are the use cases that I would be obviously showing you inside the demo as well. So, one is the diagnostic and safety. The left hand side is my mobile app that I have put the screenshot on.
[11:20] The diagnosis is happening in the this is the three the tire pressure is gone to three and the brick see understand that here tire pressure is gone to three and it will call the service center booking and accordingly send the response.
[11:33] Similarly for the hyper-personalization case, um when whenever a driver says that hey, I'm feeling a little warmer, so it understands that hey, what is the current temperature looks like and this is the dashboard or the simulated dashboard looks like which kind of
[11:46] changes the gives the output and then again gives the back the response to the user. So, cool. Uh let me quickly uh play this video. I won't be playing the full video. I'll showing you some of
[11:59] the snippets that we did specifically for this uh demo. Um >> to the Android app. >> Hope you can uh hear the audio as well, right?
[12:11] >> We'll show you the left >> So, in the left-hand side, uh I'll just show you in the left-hand side. So, maybe uh >> the left-hand >> Uh I can mute this thing. Uh so, in the left-hand side, you are seeing the uh app that I have uh screen
[12:25] casted and all the all the fields are there which fields represent the actual sensors that might be using in the car for the demo purpose. So, here I have just put on few of the vehicle and also
[12:38] uh few of the fields for the AC temperature, engine temperature, and all the four tire pressure monitoring system, right front left, front right, back left, back right, right? And ideally what would happen and there is also a driver uh interaction which is
[12:51] like who is driving uh right now. I've just written my name and based on that you can do the interaction. So, here you can see when I did the engine start, right? So, whenever you will go into the car, you start the engine, right? The
[13:03] telemetry should start flowing. So, all the data which is present inside the sensors should start flowing. In the right-hand side, you see all the uh this is a live Databricks uh dashboard which is empty as of now, but since we have
[13:17] started the ingestion via the zero bus, in the real time if I would refresh, the data is getting populated. And when I say data is getting populated, it would be taking all the week VIN number vehicle identification number, your
[13:29] registration number, who is the and all of these details of the sensors and it will getting populated inside at the right hand side of the database, right?
[13:44] And uh uh similarly uh here I have shown an example of if I change some value right in during the live session, maybe because sometimes it would happen the sensor showing different value because of some
[13:56] uh mishaps maybe, right? So, if I change the value of let's say if ideally from 90 or 95 to 125, right? And that's why how the how fast it would be refreshing inside
[14:09] your data bricks. Uh this is the dashboard and here if you can see all the uh values are getting refreshed, right? So, this is the demonstration of the zero bus. Obviously, this is a a simulated
[14:22] environment in the ideal setting it's happening directly via the uh uh the uh vehicle, right? Now, uh I want to show you uh one of the examples where I I have given about the engine is
[14:37] engine feeling hot. So, engine feel hot. Can you check, please?
[14:56] >> This is serious. Your engine temperature is critically high at 125° C. This is well above normal operating temperature. >> Right? So, basically uh when I said that I I what is the engine is feeling a little
[15:09] bit hot, then it said that hey, this is a actually a serious condition is specifically to my vehicle, which is XUV 700 if you see in the bottom, right? So, it kinds of does with full voice response saying
[15:21] that hey, you need to you These are the recommendation that you need to follow in certain standards, right? Um uh also I want to show you the car manual. So, this is for a telemetry data, but if you think about uh the your
[15:35] non-telemetry data, which may be your car manual, right? So, if I ask any question, What does my car manual say about engine overheating?
[15:58] >> Based on the XUV700 manual, here's what it says about engine overheating. If you notice engine overheating, one, immediate action is required. Reduce >> Right? So, basically when the engine temperature got froze, right? I also
[16:10] checked that hey, what's my car manual is saying about this, right? Is it okay? Is it normal or does it happen? And then it said that hey, it feel These are we feel that all of these recommendation and all of that you can do it. You can park your car for the safety and you can
[16:23] call the road roadside assistance and so on, right? And similarly now I also wanted to show you some of the insights that while taking an action
[16:35] directly to the dashboard, right? So, here if you see this is a dashboard which is taking output. So, right now if you see the temperature has been set to 23° C and the seat height is changed. And then
[16:48] if I cool things down and raise my seat, right? So, the temperature got decreased to the 18° C and the seat height also got changed.
[17:01] Right? So, uh this is how basically we can interact directly and the preferences part uh
[17:13] So, right now if you see the ambient light has been set to purple and I just wanted to share that in the lake base since you can store the long-term preferences which are session agnostic, right? If you might have set the preferred session in the two days
[17:26] back, then also it will remember it. Turn on my preferred ambient light.
[17:40] >> Ambient light has been set to orange. >> Right? So, basically if you see the change it changed the light from purple to orange and set the ambient light for it. Right? And similarly in the I can show you this is the servicing
[17:53] portal that you can interact with it. The DMS This is a servicing portal. So, from the state, the city, and the area locality, you can find out what all
[18:05] the different locations which are available and then the via the MCP your agent will actually know the live information that which which slots are available and can book a portal directly for you.
[18:17] And the last part that I wanted to show you is the since you already have the analysis >> all the
[18:31] >> So, since you have all this telemetry data, we are kind of using inside the post talk analysis to understand like what is the behavior of the data, what is the weekly distance kilometer, right? What is the speed distribution? And that is getting used by the vehicle's data
[18:43] because they already had a different finance firm with them. It's a conglomerate, so which can be used to have the insurance assessment whether this guy is a medium risk or he does a rash driving or so and accordingly you can plan out the
[18:56] insurance or the servicing as per the need. So, yeah. This is the whole architecture we are kind of building in the for the brick city where we are incorporating each bits and parts right
[19:09] from the data ingestion to the to the speech to text text to speech your Bixby being the supervisor agent interacting with the memory interacting with the vector search interacting with your tools and gives you the response on
[19:23] the fly. So, that's all I had. Thank you for attending.

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.