The Garage Podcast : S3 EP16
Neel Mitra of AWS
Neel Mitra, the Worldwide Solutions Architecture Leader for Data and AI at AWS, discusses the evolution and applications of AI in the automotive industry. He highlights the transition from classic AI to large language models and agentic AI, emphasizing their potential to enhance vehicle diagnostics and performance. With over twenty-one years of experience, Neel shares insights on the importance of data integration and collaboration in optimizing automotive technology. He notes the growing trend of software-defined vehicles and the need for continuous learning through MLOps. The conversation underscores the significance of AI in transforming automotive systems and the benefits of these technologies. The conversation touches on Sonatus AI innovations such as Sonatus AI Director, a new solution for in-vehicle edge AI, and Sonatus AI Technician, which uses LLM's to provide a better diagnostic experience.
Listen to audio only version:
Episode Transcript | Neel Mitra of AWS
Table of Contents
0:00 Introduction to Neel Mitra and AI Master Class
Today in The Garage, our guest is Neel Mitra, worldwide solutions architecture leader for data and AI at AWS.
In today’s master class on artificial intelligence, we start with the origins of AI. What was classic AI? What is a large language model? What is agentic AI?
And then we talk about automotive.
How does AI play in automotive today? And what are the opportunities to do new and incredible things with AI and automotive in the future? You’re really gonna enjoy today’s conversation.
Let’s go.
Welcome to The Garage. I’m John Heinlein, Chief Marketing Officer with Sonatus. We’re here today with Neel Mitra from AWS. Neel, welcome to the garage.
Thank you.
Neel , we’ve been working together for a number of years. It’s so exciting, and we were finally able to get you to come visit us and record an episode with us. Start by telling us about you and your background.
Yeah. Thanks for having me. It’s a pleasure. We had been working on so many different initiatives, and finally, the day is here. So I grew up in India, had been in the industry for twenty one years. Before that, I completed my studies in engineering, masters from India, worked there for a few years in consulting, then came to U.S.
And then, you know, since then, I have worked with a lot of Fortune Five Hundred customers in fintech and health care, and
1:26 Meet Neel Mitra
I never knew I will work on automotive until 2016 when I joined Amazon Web Services.
And those days, you know, cloud was so new. I remember I used to go to customers, and then the customers will say, hey. You guys, sell books. Right? What are you doing here?
So I will tell them, like, what we do in the enterprise space, and then I’ll tell them what we’re doing in the infrastructure. At that time, we’re very heavy in infrastructure.
And as you know, AWS had been the pioneer in cloud computing, and we had been, like, you know, the top of the gardener for infrastructure even today, I think. So, yeah, since then, it’s nine years. Like, you know, days fly. And at AWS, we like to say, like, every day is day one. So, again, thanks for having me. And, it was great learning a lot, you know, from all of you as well as we were in the journey in the automotive space, which we’ll dive deep into.
Thank you so much. And we always like to start with a fun fact about our guests. You have to tell us a fun fact about you.
2:24 Yoga, Hot and Cold
That’s a tight spot. Okay.
So, you know, a lot of things in my life like happened for no reason, I believe, like even joining AWS and working on automotive, I don’t know why that happened, but it turned out to be real fun.
Same way yoga entered my life. And I don’t know why and how, because I never heard about yoga in India and I grew up there, right? But, you know, there was a point where I was traveling a lot and it was too much going on and I felt like, you know, there is something has to be there which takes care of me holistically, Not just physical, but mental and energetic and everything. And I tried this yoga class and I liked it, and then it was on and off.
I had been doing that probably for eight, nine years now. But last year, again, suddenly, I don’t know how random it is, but I thought I want to be a yoga teacher as a hobby. And I started taking a certification. OEMs.
That’s scary. My wife also tells me, why did you do that? But, yeah, that’s a fun fact. Probably in the last one year, I completed my yoga trainer certification, two hundred hours.
Now I’m doing the masters, which is three hundred hours more.
So it’d be fun, probably my retirement plans to be a yoga teacher.
Fantastic. And what style of yoga?
Vinyasa.
Fantastic.
My fun fact I always try to think of a fun fact for our guest. My fun fact is quite some years ago, maybe about ten years ago or so, I, for a couple years, was very devoted to Bikram yoga.
Oh, nice.
Which if if our listeners are not familiar with it, it’s when you have one hundred and five degree room. Yeah. Hot yoga. And you got high yoga as they call it, and you get in the room and you do yoga for ninety minutes. It’s a very specific ninety minute, sequence. And by the end, like everybody is soaking wet.
I did one of those in New York. I, I went to so many yoga places and finally I’ve settled into one or two, but, yeah, it has been fun.
And yoga is fantastic.
Of course, the strength, of course, the flexibility, but also the mindfulness and the kind of calm Yeah. It gives you after ninety minutes with a hot room, but just in general, you you leave feeling so, peaceful, and I think it gives you great ideas. I had some of my best ideas at minute forty five, maybe.
And, you know, more practically, I know we’ll talk a lot about AI later today, but more practically, I see when machines are doing so much work for you today. Right? Yeah. If we don’t be mindful and operate a higher level of frequency as humans, it’s going to be difficult.
Like, what are you going to do? You’re writing code, machines are doing it. Right? So that’s where I feel like my yoga might be useful to me as well from a mindfulness standpoint.
We’ll see how that goes.
Alright. So tell us about your role at AWS. What are you working on these days?
Yeah. So, you know, like, it’s all over the place. Like, I had been focusing a lot on automotive, as you guys know, for the last few years. So, had been working with a lot of connected mobility starting with then moved into software-defined vehicles, working with a lot of, you know, worldwide customers and partners, you know, starting from, you know, the Boschs and the DENSOs. And a lot of customers I cannot name, but, there are the BMWs and others as well, which I had been very busy with.
But in the last one year, I have shifted my focus to be more horizontal. That means I’m working with a lot of automotive manufacturing and enterprise customers on a wide variety of use cases. Those are not like super niche automotive, which I did in my architecture and engineering roles last few years. But it has been fun because, you know, there is so much changes happening in the machine learning space with AI and agents and everything.
The opportunities are endless and it’s not just about one industry that I’m seeing. So I’m trying to learn more. And that’s the best thing I feel like being at AWS. I feel like I am being paid for learning so much. Like if you’re in a university, but you are not paying. You are being paid and you are able to work with so many talented people across the globe and doing something to make this world a better place.
So, yeah. So that’s my journey in the last one year, doing a lot of work in the AI space with OEMs and RAGs and Agentic AIs for different set of use cases, including manufacturing and automotive.
Great.
Yeah.
6:25 Introduction to Artificial Intelligence (AI)
So you mentioned AI and and a lot of times we talk about vehicle and vehicle software, and we’ll get to that a little later. But this episode, we’re gonna talk a lot more about AI than usual. Okay. So let’s let’s talk about that first from a high level perspective because I I think our listeners have probably a range of exposure to these kinds of technologies.
Yeah. And I think when most people hear AI today, they probably think ChatGPT or, you know, search and things like that. Yeah. But that’s just a a tiny fraction
of the potential for AI.
So I wonder if you could give us with a twenty-thousand-foot level first. Yeah. What’s the spectrum of AI algorithms and AI technologies that we could think about?
Yeah. You know, a fun fun fact before I get into that, I watched the movie Terminator, like fifteen years back, probably the first time. Right? And I was really scared, like, what AI can do for you, like, how you’re fighting against the AI and stuff like that.
But then when I actually started working on it, I felt like, okay, it’s like probably a few hundred years away. Even if that happens. I don’t know. Like, it’s a science fiction.
Right? We’ll we’ll figure it out. But you are absolutely right. Like, there is so much that is happening in this space.
Like, AI is not new. Right? Those research papers were from 1930s and 40s and 50s.
And then, you know, we at that time, we started talking about, like, supervised model. That means you are giving some labeled data. You are trying to make the machine smarter so they can do some of the human stuff.
And then as it evolved and evolved and evolved, you saw like the introduction of unsupervised models. So it’s not just labeled data. It could be unlabeled data. It could be multimodal data. It could doesn’t have to be just text. Right? It could be images and videos and audios and whatnot.
And believe me or not, like in the nineties when AI was not considered so cool, Amazon used AI at that time.
So if you went to that first Amazon site ever, I didn’t, I don’t know if you did, it used to give you an option that, hey, if you like a specific genre—A recommender system—Yeah, recommendation system. Right. And that was AI. That was the early nineties. Then Google did it, like with this web search and stuff.
Netflix used to do that in the early days when Netflix was sending you a DVD and something like that.
Exactly. So AI is not new, and this kind of supervised model and a little bit of unsupervised model had been there for quite a long time.
And even if I remember the early days when I was in, let’s let’s say, fintech industry, you were doing a lot of this fraud detection, anomaly detection. That’s all kind of AI. Right?
But then when I got into automotive, the first set of AI use cases I saw is like the cruise control, which is essentially you are, you know, getting all this multimodal data and you are figuring it out, keep the car on lane.
There are lane departure kind of use cases, right, keeping the lane in in control and stuff like that. So those are all traditional machine learning models. So back to your question, yes, there are supervised and unsupervised and deep learning came into the play where, like, how do we think like humans and neural networks and all those cognitive functions we started to slowly put into our vehicles. And the vehicle started doing the cruise control. And then we are talking about different levels of autonomy these days.
And there are so many use cases, but from there as we are evolving, I think attention is what you all need. I think that is the name of the paper. I forgot the exact name, but that came out from Google in, like, in 2018-2019. Right? Attention is what you need.
So that talks about the self attention mechanism.
And if you think about the old AI, which is, like, the neural network CNNs, convolutional networks, and the RNNs, recording networks, You will see the problem was you have all these words, right? But the AI couldn’t figure out what word or what token will come next. It is very limited. Like if I say, Hey, John bought a Rivian or Neel bought a Tesla or whatever it is, right?
The AI couldn’t figure out a long sentence. It can probably go only few words.
But with the self attention mechanism, with transformer architecture that got invented in 2018-ish, that’s changed. Like now you can have thousands and hundreds of thousands and billions of tokens like this. And with the self attention mechanism, AI with the encoder decoder architecture can figure out where the next word could be. So this contextual awareness and the ability to process humongous amount of tokens together changed everything. That’s the inception of large language models, which as you mentioned is ChatGPT everyone was talking about, but those were all the foundation stepping stones that led to LLMs and there is more to come.
Yeah.
11:01 What are Large Language Models (LLMs)?
So now that you’ve just mentioned on, I think what is the next level and this key innovation of these transformer models and the T in GPT is transformer model.
The tell us about how this transformer models and maybe LLMs more generally, how does that change the game? What are some of the things that are now possible with transformer models?
Yeah. I think the biggest innovation we have seen in is content generation. Right? So you are in marketing yourself, so I’m sure you have seen so much productivity, right? You can create a PowerPoint and you can create all these stories and videos and audio.
Like, you can create a podcast, like, I think Google has one, right, where it sounds like two humans talking. This one’s not AI generated!
This one is not AI generated, absolutely, these are real humans talking. But the content generation, I think one of the, like, most innovative thing that came out of all this ChatGPT stuff.
And then we started learning about seventy billion tokens being used to train this model or one hundred billion tokens to train this model. But then you raise a very valid point. How does it help the broader industry? That’s where we see a lot of developer productivity suites coming out.
So for example, with Amazon, we have Q Developer, or we recently launched something called Kiro, which is agentic IDE that is using the power of the large language models to make all those decisions. So it’s not only content generation for marketing or for storytelling. It could be content generation like the code you are writing, and that could can be anything. It could be a Java or Python code, or that could be a code that is in mainframe, which is fifty years old.
And a lot of our customers, believe it or not, especially in the fintech, healthcare, those kind of industries have so much mainframe presence and they cannot move it, right? So now if AI is able to understand all this code and able to refactor the code and bring it to the cloud and make it more modern, that’s a huge ROI for the customers.
Yeah. And I think what’s interesting about about, LLMs in general is you’re seeing now instead of thinking about a general purpose LLM and people talk about sort of AGI or artificial general intelligence, which is we’re not there yet. In fact, what actually happens is these these LLMs are being trained for specific functions, image creation or audio manipulation or or whatever. And as a result, they can be a kind of an expert in that discipline.
Over time, those models will converge. You’ll have more kind of a consolidated expert. Yep. But today, you’re seeing these special expert models that are allowed to do able to do these incredible things you mentioned.
Absolutely. You’re spot on. And that’s where I think this industry will evolve so much.
And we have a LLM marketplace at AWS as well. Right? So for a lot of the domain specific functions, if you think about even, like, predictive maintenance, which had been a big problem in the automotive industry for such a long time. And if or, you know, a lot of time series forecasting models, which had been, you know, a need for a lot of different kind of use cases.
If you think about large language models that might be trained focused on those kind of domain-oriented problems, that’s a game changer. Yeah. And that’s where we launched something called distillation last year. So it’s not about, you know, going for the big, big, big, big, big, but going for the data quality that truly matters for your line of business.
And that could be automotive. That could be health care. That could be fintech. But you’re spot on.
I think that’s where the industry is moving, and we’ll see a plethora of models. And then the customer choose the right model for the right job.
Yeah. An example that many people know, but I think is really compelling to me is, when you think about, medical diagnostics.
Yep.
X-rays in particular. If you ever had an X-ray, you know, you have an X-ray and then they sort of the the technician sees the X-ray live, but they’re not allowed they’re not allowed to tell you. But they, like, see hundreds and hundreds of X rays. And then it goes to a doctor who sees tens of thousands of X rays, and he says, oh, you have a broken bone or you don’t have a broken bone or you have this or you have that.
Right. But they’ve trained models on tens, hundreds, thousands, millions of X-rays, and these models are able to do as well or, in in some cases, better than a doctor because of the sheer volume of images they’ve seen, and they can see things that doctors can’t see. Now I’m not saying that they’re gonna replace doctors, but what it can do potentially is do some filtering. It’s saying this is the five percent of X-rays to look at.
The other ones is very clear, for example. Yep. Whether that reduces cost or improves patient outcome or reduces turnaround time or whatever.
The point is these are new tools. Yeah.
And they can be used in new ways. Yeah.
15:24 What is Agentic AI?
Now you mentioned, earlier on another word, which is agentic. And I think I bet most people don’t yet fully appreciate what agentic means.
It’s a very important emerging trend, but it’s also early in the life cycle. Can you talk about what agentic means? Yeah. What is some of the opportunities that agentic AI can bring, but also what are some of the challenges and limitations today?
Yeah, absolutely. So I think there was a saying that last year or the last few years were the year of OEMs, because you see LLMs like being launched by all these different providers almost every week. And then came the RAG, the Retrieval Augmented Generation, all the era of the chatbots. Everyone is building chatbots using these LLMs.
But this year is supposedly the year of agents. Now why agents? Because you need to take action. Like,
even as humans, right, if we continuously think and do not take any action, we’ll go nowhere.
So similarly with all these OEMs and chatbots, you can query all this data, and that might be a good start. But then to execute on top of that data, you need certain functions. Now in traditionally, how do we do that? We use microservices.
We use, like, so like, SOA architectures, REST APIs, and and stuff like that. Right? But in the AI world, we are talking about agents because agents have the decision making power, and they could be autonomous. Now what level of autonomy?
That depends on the use case, and that’s a growth area where we’ll learn a lot more. For example, you know, in the automotive, we had been talking about fully autonomous vehicle for such a long time, but because of different complexities, that is not yet possible. The same thing will happen with a lot of this agentic workloads as well. We can keep talking about fully autonomous agents doing everything, but that might have lot of trade offs when they’re making the decisions.
Now you’re already seeing examples of, something as mundane as say calendar scheduling or find me a flight from here to there in this time frame at a good price. And so and then, obviously, as as time goes on, I think we, society, and then, of course, every individual may be in a different place, may be more willing to give more authority, more, agency—pun intended—to the AI to do things for them. Or other people could sort of have it do the work and then check back, does this look okay for you before they execute? And I think we’ll see all of those above. But I think the key point, though, is is by by having the loop of not just analytics of getting an answer, data analytics, but also taking some action, potentially going towards the loop closing, you begin to get the ability to have, some feedback cycle that can improve the capability and throughput of AI and potential.
Yeah. The biggest differentiator is when we write software to date. Right? It is mostly rule based.
So you are essentially saying, hey, a microservice or API, do x y z. And to do that, that will take an input data, and that will send an output data. And you have to codify that logic.
But with agents, because it has the power of the large language models, it is able to make so many decisions. So you may not have to be explicit all the time, but the trade off there is, is the power of thinking. Like even as humans, right? If we keep on thinking too much, we’re going to exhaust us.
We are going to lose a lot of energy. We need to drink more. We need to eat more. So same thing for the agents.
If it is thinking continuously, and it is kind of non deterministic in nature, you are paying the cost because the agent is running for a long time using a lot of memory, a lot of CPUs and stuff like that. So long story short, I think you know, agent is the future, of course, but if there are logic based workloads, like if-then-else straightforward, you still go to microservices. You still build those microservices.
Whatever their decision making needed, like a supply chain kind of use cases, which is a big challenge today, right, especially with all these tariffs and stuff like that. There’s so much uncertainty.
And if you want this agent to figure out the inventory, you have to figure out, like, what is the vendor status, you have to figure out the logistics and transportation, Those are the great use case where you cannot codify a lot of the logic. The agents or the multi agentic orchestration need to go to different systems to take care of different actions. So those are the great use cases for agents, and, yes, that is the future.
But don’t try to put all the eggs in one basket and do agents for everything if, you know, it can be fit by a rule based logic.
Right.
So in my opinion, both these worlds will continue to stay, which is more deterministic and nondeterministic and use the right tool for the right job.
19:56 The evolution of AI
So if I if I sort of pull back and I summarize everything you’ve just said, what I hear is AI is growing. It’s getting more capable. Yeah. But one shouldn’t think of it as a monolithic thing, but there’s a spectrum from simpler models to to more complicated models to potentially more, independent, obviously autonomous. I don’t wanna confuse the meaning of that, but more, agentic, action oriented models.
And all of those will continue to be valuable in the future.
Yep.
And the mix probably will evolve over time.
Is that a way to think about it?
Absolutely. And the same thing if you think about our computer, storage, and database, everything happened like that. Like we started from, let’s say, all these monolithic mainframe OEMs. Right?
And then we moved into decentralized systems. We started talking about VMwares and lot of this, and now containers. And now we talk about serverless functions, but all those big physical machines still exist. But at the same time you have the serverless function, which is a single piece of code running on point five gig of memory.
So this whole plethora of capabilities in the compute list still continues to stay. Databases. You have all these giant monolithic databases, right? But you also have small databases like time series databases.
You have other type of data, NoSQL databases.
So the same thing will happen in the AI space where you have a plethora of LLMs . And we are here. You know? That’s what I joke. Like, as solution architects, that’s sort of our job to work backwards from the end customer to figure out what tool
is right for your use case, because don’t try to fit in everything in one basket. It’s the right tool for the right job.
Great.
21:24 AI applications in automotive
So we have to talk about automotive. Sonatus, and our podcast is definitely about vehicle technology and vehicle software. So let’s talk you mentioned a few things, but let’s talk about some ways that AI is being used in vehicles today, and then we can talk about kinda where it’s going in the future.
Yeah. So, you know, in the last ten years or so, as I was in the automotive industry, I’ve seen so many use cases. Like, my background was in big data and IoT. And in addition to automotive, I was also working with a lot of connected home kind of products.
Right? And that’s why connected mobility and all those connected home kind of were similar. You were getting a lot of those telemetry data. For automotive, you’re getting primarily from the telematics unit.
And they were doing all the fleet management and things like that.
Then I started working on the OTA side of the world because we started talking about, hey, like automotive should be also in over their upgrades and not just the telemedics units or the infotainment unit, but how can we make the entire automotive side of things more upgradable? And it has to be upgraded over the time and the cost doesn’t depreciate.
And that’s where the concept of software-defined vehicle started emerging. And it was not just about updating over the air.
As you guys know, you have been, you know, kind of a SME and expert in the space. It’s also about transforming the architecture within the vehicle because if you have hundreds of ECUs out there, it’s really, really difficult. So how you can bring that ECU consolidation and make it more zonal-based architecture or domain-based architecture, and you can push all of this, you know, software updates or collect a lot of the data from the vehicles to bring the next generation of experience.
So I think for AI, it has been always there as we’ve mentioned. Right? Different forms of ML had always been there. When I was working with one of our partners, BlackBerry, a few years back, we’re talking about synthetic sensors as well, where we deploy a lot of the sensors, which are machine learning based, or it could be logic based.
And you are collecting all this data, processing the noisy data on the vehicle before you do something on the cloud because sending the data in and out is also cost prohibitive at many times. So, sorry for the long winded answer, but I believe that with generative AI and agents, the entire supply chain to connect, connected mobility
or collection of the data and how much high quality data you can actually send to the cloud. How much work can you do, which are non deterministic in nature on the vehicle itself with the EE architecture and software defined vehicle, is going to be crucial.
And that’s where your earlier point of having the right model to do the right job.
You don’t have to have the hundred billion token model in the car. You can probably have, you know, a ten million token on the car, which is good enough for what you are trying to do. And that’s where all this great softwares that you guys are building, which is focused on the vehicle software, I think would be immensely valuable.
24:11 Sonatus AI Technician
Yeah. Thanks so much. Appreciate that. And let’s let’s get into that. I mean, I I think as we mentioned at the beginning, I think people tend to think, and, there’s a perception that AI in vehicles is only autonomous driving or it’s only automatic braking or it’s only, adaptive lane keeping.
And those are phenomenal. I’m a huge fan of all those technologies, and I love them. Yeah. But what what we’re trying to do at Sonatus, and I think that there’s a missed opportunity and we’re trying to raise awareness, is I think that there’s much more that’s possible.
So I’ll give you a couple of examples. And we’ve been collaborating with AWS for years now. We really love your partnership.
Last year at the Consumer Electronics Show, that is to say January twenty twenty five Yeah.
We showed off a new product called our AI Technician, and this was running on top of AWS, using large language models. And what we showed was that there’s an opportunity to use, LLMs and use analytics to do a better job of understanding problems in the vehicle
And then leveraging the things like the the software, the, the data infrastructure that Sonatus provides for better, finer grain data Yeah.
To be able to identify, hey. Is this a is this a problem? Isn’t it a problem? Is it an urgent problem?
Is it a non-urgent problem? Is it a fleet-wide problem? Yeah. Is it a one-vehicle problem?
So this AI Technician idea has gained, has garnered a huge amount of interest. And we we also use that we talked about this agentic idea that it wasn’t one source of data. It’s not just, okay, we just look at the the, the trouble codes, DTC codes, diagnostic trouble codes. We’re not just looking at DTC codes and putting it into a RAG or putting it into LM.
Yeah. But rather—and this was the point you made earlier—we’re pulling together data from many sources. And combining the data from many sources to get a better answer than any one source alone. And I think that’s where a power that AI brings.
That’s that’s going to be hugely powerful. I agree with you. Like, we call it data center on wheels, And it truly is because as you were saying, there are like so many network topologies, right? There is CAN, there is ethernet, there is, like so many others out there.
And what we had to struggle for a long, long time is how to deserialize lot of these different type of signals into something standard. And that’s why we started working on VSS specifications from COVESA, but not every OEM uses like VSS, right? Everyone has their own thing. So now how do you standardize all this data coming from all this network topologies, different kind of ECUs, right, and then making sense out of it in near-real time to bring that personalized experience or even predictive maintenance that your brake pad is having wear and tear and needs to be replaced, send it to the dealership.
And then you go to the dealership. Dealership says that, no, it’s not a problem. We’re going to fix it. Right?
So all of this is this holistic value chain. I think that AI can solve it. Not just, as you said, like the fancy infotainment system. Sure.
Like I worked on the Alexa SDK when it first launched for embedded devices. And I still remember, like we did a reinvention with NXP and we showed customers how you can embed, right? Alexa, what stuff. So all this infotainment kind of stuff we had been doing for such a long time and ML had been there for such a long time, but there is so much more to do with the AI of today’s age, as you’re saying, because it’s the next evolution, and it’s everywhere.
27:28 Sonatus AI Director: In-vehicle Edge AI
It’s so I’m so glad you said that.
And that last point is a perfect transition to the the last point we want to make as we as we’re sitting here today, we’re so excited that Sonatus has just launched our newest AI product called Sonatus AI Director. And and what I think that building even on what we talked about a moment ago with AI technician, our recognition is that the AI today tends to be centered in ADAS units, generally speaking, or, obviously, in certain cases, chatbots in the IVI unit to do things like where’s the nearest Starbucks, or am I gonna hit traffic on the way home? Yeah. But the reality is that’s a tiny fraction of the vehicle infrastructure. It’s a tiny fraction of the vehicle subsystems.
And so our recognition was there’s a huge untapped potential to use AI in more vehicle subsystems for a range of applications, whether it’s better diagnostics, better preventative maintenance, optimization, efficiency improvements, tuning over time, and so on. So what AI Director did, and this is, AWS is one of our launch partners, and so we’re so happy to have you, help us kick this off, is showing how we, can provide an infrastructure to accelerate the deployment of AI into the vehicle, not in the cloud. The cloud is important. The cloud will always be important.
Yep. But you there are certain applications that you have to do locally, whether it’s latency sensitive or there’s too much data to send to the cloud or there’s privacy concerns, or there’s IP- protection concerns. How can we bring those models, where appropriate, closer to the vehicle running in the infrastructure of the vehicle and solving problems, cost reducing, providing efficiency.
So we were so excited to launch that recently and so happy to have you. You’ve been looking at this kind of problem for a while. Do you see this as a real need in the industry?
Yeah. I absolutely do because, you know, data is still the differentiator.
Because we keep on talking about AI, but, sorry for my language, but it’s going to be garbage in, garbage out if you do not have the right quality of data. And that’s what I do believe. You know? Like, we still need to obsess about having high data quality, high data governance, as you mentioned, all these different kind of optimizations.
And then your AI, whether it’s LLM based AI, whether it’s agentic based AI, is going to generate that ROI for you. If you just think you will get this LLM, which is trained on Internet data, and it will do magic for your business, it’s not going to. I mean, that’s the reality.
And that’s where a lot of the solutions that, you know, you guys are pioneering on, I believe it would be, you know, super useful for the customers.
Yeah. I thank you for that. And and one of the things that occurred to us, and we’ve validated this and it continues to be the case is there are so many sector expertise. There’s so much knowledge and know how for batteries, or propulsion or safety or tires or whatever.
30:25 Leveraging sector expertise in AI and data exchange
That that kind of expertise exists out there. But the challenge today is OEMs have to integrate each of these models individually at high integration cost. How could we make a more common infrastructure for them to draw upon innovation from these industry experts in a way that’s very scalable for them? And that’s what we are trying to do. And so far, there’s been very, very positive reception for that.
Yeah, and this is amazing because even if you think about this multimodal OEMs Right. They are actually bringing a lot of the knowledge, although they’re still internal data. But more and more, it is being used by different, you know, industry verticals. And similar to how, you know, like, ten years back, I had seen, like, lot of customers are actually contributing something called data exchange.
That means they have the trader, which is high quality data maybe for finance industry or health care or auto, and they’re sharing it either in open source model or commercial model, whatever it is. And that can be used by different other verticals. I believe same thing will probably happen for this space as well. And when this happens, the LLMs are so powerful.
You use that as your brain and bring all this best practices. As you said, like EV range anxiety is such a big problem even today.
Right.
I talked to so many, OEMs even, and they talk and they share the same problem. Like there is EV range anxiety and what can we do? And this cannot be optimized until you have this entire data flywheel of collecting the data from the tires, collecting the data from different ECUs and then consolidating it and doing that predictive or preventive kind of maintenance, even range optimization.
It seems simpler, but it is not because we’re also putting so much capability within the car. Like my kid wants to watch Netflix while for a long drive. Right? And if you’re putting so much battery out of the system and still expecting your range to be five hundred miles, that’s not going to happen. So how do you optimize all of this? And the answer is data and AI on top of it, and you continuously learn, fine tune your model, whether it’s a traditional model or a large language model. Right?
And then that is the MLOps that, you know, you can do.
32:30 Anticipating the future, today
Exactly. And you mentioned range anxiety. One of the the sources from that is also the battery. Vehicle battery technology, the understanding of that is evolving rapidly.
One of our launch partners is a battery management system, battery-health leader called Qnovo. Mhmm. And we were happy to have them as part of our launch.
And what they’re looking at is they have in incredible innovation in looking for failures in battery, looking for the ability to charge batteries faster and so on, but you have to have the data. You have to have the model. And so one of the things we’re doing with them is helping use the same infrastructure to provide a way to land that model. So that’s a subsystem you wouldn’t have thought of, but it can leverage the same infrastructure.
Exactly.
So and then the other point you made, which I think is really important, is that we as an industry and we as technologists, our mission always is to anticipate the future. And and sometimes anticipating the future means that you know there’ll be things in the future that you don’t know.
Yeah.
So how do you put in place the framework so that when the next innovation comes, you can use it?
You can avail yourself of it. Because too often, with many technology, but certainly with vehicles historically, is you design it because it’s a car, it drives, and that’s all I need. But then something comes that wasn’t anticipated, and your vehicle becomes obsolete or less competitive in the marketplace or resale value goes down. How do we put in place the capability to extend vehicles over time for range optimization, for new services, for new safety technologies after shipment. And that’s one of the extensibility benefits we’re trying to bring, and and we feel like there’s a huge need in the industry.
34:04 Using the cloud for digital twin
Yeah, yeah, absolutely. And this is where, you know, I still believe the, you know, power of cloud is going to be so important. Right?
I have not directly worked on some of these initiatives, but my peers have like the virtual ECU on the cloud. Right? Where you were bringing a lot of this NVIDIA stuff that you were probably using on the car. How can you keep running them on the cloud and do all these simulations?
And especially if you’re going to that EE architecture where you have consolidated ECUs, you have the car on the cloud. And then you are running end-to-end simulations on all of this. And eventually, you are figuring it out, like, what makes sense at this point of time? Because a car is a different animal.
It’s not like a headphone, which is a smart headphone. Right? You know, like, I don’t want to preach to the choir. Like, it’s like safety critical systems that have deterministic actions you need to take.
The airbags getting deployed. You need to take an action immediately. This is not like any other connected products. This is the most complex product out there.
35:03 Conclusion
So I think, as you said, like it’s continuous learning. We will be accelerated on the cloud and then partner on the edge side, but we need all this data. And then machine learning and OEMs need all this data. We need all those MLOps. We need all the SDLC best practices. That’s not going anywhere. It’s just the tools changing.
And we need to use the right thing for the right job.
It’s a perfect summary, you said, because while we talked for a moment ago about the vehicle edge, the point is really that you need all of the above. Yep. You need strong cloud. You need strong data.
You need strong services. Yep. And you need data in the vehicle, and they need to work closely together. And that’s really the story.
And that’s why we partner with you. Yeah.
Yeah.
And that’s why we’re so happy to have you here. Likewise. It’s such a joy to talk about this, and it’s so exciting to talk about AI in more detail. We’re gonna be talking more about AI in in future episodes. But thank you for the master class, really, in all of these technologies.
Thank you for having me. Really appreciate it, John. It was so fun chatting with you.
If you like what you’re seeing in this episode, please like and subscribe to see more like it. We appreciate you joining us and hope to see you in another episode very soon.
Recent Episodes
The Garage Podcast
Marc Fecker of Robo
The Garage Podcast
Suat Kusefoglu of Capgemini
