(bright music) >> Hey, welcome back everyone, theCUBE coverage here at AWS re:Mars 2022. I'm John, your host of the theCUBE. re:Mars, part of the three re big events, re:Invent is the big one, re:Inforce the security, re:MARS is the confluence of industrial space, of automation, robotics and machine learning. Got a great guest here, Muhammad Faisal senior consultant solutions architect at Capgemini. Welcome to theCUBE. Thanks for coming on. >> Thank you. >> So we, you just we're hearing the classes we had with the professor from Okta ML from Washington. So he's in the weeds on machine learning. He's down getting dirty with all the hardcore, uncoupling it from hardware. Machine learning has gone really super nova in the past couple years. And this show points to the tipping point where machine learning's driving space, it's driving robotics industrial edge at unprecedented rates. So it's kind of moving from the old I don't want to say old, couple years ago and the legacy AI, I mean, old school AI is kind of the same new school with a twist it's just modernized and has faster, cheaper, smaller chips. >> Yeah. I mean, but there is a change also in the way it's working. So you had the classical AI, where you are detecting something and then you're making an action. You are perceiving something, making an action, you're detecting something, and you're assuming something that has been perceived. But now we are moving towards more deeper learning, deep. So AI, where you have to train your model to do things or to detect things and hope that it will work. And there's like, of course, a lot of research going on into explainable AI to help facilitate that. But that's where the challenges come into play. >> Well, Muhammad , first let's take, what do you do over there? Talk about your role specifically. You're doing a lot of student architecting around AI machine learning. What's your role? What's your focus. >> Yeah. So we basically are working in automotive to help OEMs and tier-one suppliers validate ADAS functions that they're working on. So advanced driving assistance systems, there are many levels that are, are when we talk about it. So it can be something simple, like, you know, a blind spot detection, just a warning function. And it goes all the way. So SAE so- >> So there's like the easy stuff and then the hard stuff. >> Muhammad : Exactly. >> Yeah. >> That's what you're getting at. >> Yeah. Yeah. And, and the easy stuff you can test validate quite easily because if you get it wrong. >> Yeah. >> The impact is not that high. The complicated stuff, if you have it wrong, then that can be very dangerous. (John laughs) >> Well, I got to say the automotive one was one was that are so fascinating because it's been so archaic and just in the past recent years, and Tesla's the poster child for this. You see that you go, oh my God, I love that car. I want to have a software driven car. And it's amazing. And I don't get a Tesla on now because that's, it's more like I should have gotten it earlier. Now I'm going to just hold my ground. >> Everyone has- >> Everyone's got it in Palo Alto. I'm not going to get another car, no way. So, but you're starting to see a lot of the other manufacturers, just in the past five years, they're leveling up. It may not be as cool and sexy as the Tesla, but it's, they're there. And so what are they dealing with when they talk about data and AI? What's the, what's some of the challenges that you're seeing that they're grappling with in terms of getting things integrated, developing pipelines, R and D, they wrangling data. Take us through some of the things. >> Muhammad: I mean, like when I think about the challenges that autonomous or the automakers are facing, I can think of three big ones. So first, is the amount of data they need to do their training. And more importantly, the validation. So we are talking about petabytes or hundred of petabytes of data that has to be analyzed, validated, annotated. So labeling to create gen, ground truth processed, reprocessed many times with every creation of a new software. So that is a lot of data, a lot of computational power. And you need to ensure that all of the processing, all of handling of the data allows you complete transparency of what is happening to the data, as well as complete traceability. So your, for home allocations, so approval process for these functions so that they can be released in cars that can be used on public roads. You need to have traceability. Like you can, you are supposed to be able to reproduce the data to validate your work that was done. So you can, >> John: Yeah >> Like, prove that your function is successful or working as expected. So this, the big data is the first challenge. I see that all the automotive makers are tackling. The second big one I see is understanding how much testing is enough. So with AI or with classical approach, you have certain requirements, how a function is supposed to work. You can test that with some test cases based on your architecture, and you have a successful or failed result. With deep learning, it gets more complicated. >> John: What are they doing with deep learning? Give an example of some of things. >> I mean, so you are, you need to then start thinking about statistics that I will test enough data with like a failure rate of potentially like 0.0, 0.1%. How much data do I need to test to make sure that I am achieving that rate. So then we are talking about, in terms of statistics, which requires a lot of data, because the failure rate that we want to have is so low. And it's not only like, failure in terms of that something is always detected, and if it's there, but it's also having like, a low false positive rate. So you are only detecting objects which are there and not like, phantom objects. >> What's some of the trends you're seeing across the client base, in terms of the patterns that they're all kind of, what, where's the state of their mindset and position with AI and some of the work they're doing, are they feeling, you feel like they're all crossed over across the chasm so to speak, in terms of executing, are they still in experimental mode in driving with the full capabilities is conservative or is it progressive? >> Muhammad: I mean, it's a mixture of both. So I'm in German automotive where I'm from, there is for functions, which are more complicated ones. There's definitely hesitancy to release them too early in the car, unless we are sure that they are safe. But of course, for functions which are assisting the drivers everyday usage they are widely available. Like one of the things like, so when we talk about this complex function. >> John: Highly available or available? >> Muhammad: I would say highly available. >> Higher? Is that higher availability and highly available. >> Okay. Yeah. (both laughing) >> Yeah, so. >> I know there's a distinction. >> Yeah. I mean >> I bring up as a joke cuz of the Jedi contract. (Muhammad laughs) >> I mean, in like, our architecture. So when we are developing our solution, high availability is one of our requirements. It is highly available, but the ADAS functions are now available in more and more cars. >> John: Well, latency, man. I mean, it's kind of a joke of storage, but it's a storage joke, but you know, it's latency, you got it, okay. (Muhammad laughs) But these are decisions that have to be made. >> Muhammad: They... >> I mean. >> Muhammad: I mean, they are still being made. >> So I mean, we are... >> John: Good. >> We haven't reached like, level five, which is the highest level of autonomous driving yet on public roads. >> John: That's hard. That's hard to do. >> Yeah. And I mean, the biggest difference, like, as you go above these levels is in terms of availability. So are they these functions? >> John: Yeah. >> Can they handle all possible scenarios or are they only available in certain scenarios? And of course the responsibility. So, it's, in the end, so with Tesla, you would be like, if you had a one you would be the person who is in control or responsible to monitor it. >> John: Yeah. But as we go >> John: Actually the reason I don't have a Tesla all my family would want one. I don't want to get anyone a Tesla. >> But I mean, but that's the sort the liabilities is currently on you, if like, you're not monitoring. >> Allright, so, talk about AWS, the relationship that Capgemini has with AWS, obviously, the partnerships there, you're here and this show is really a commitment to, this is a future to me, this is the future. >> Muhammad: Yeah. >> This is it. All right here, industrial, innovation's going to come massive. Back-office cloud, done deal. Data centers, hybrid somewhat multi-cloud, I guess. But hybrid is a steady state in the back-office cloud, game over. >> Muhammad: Yeah. >> Amazon, Azure, Google, Alibaba done. So super clouds underneath. Great. This is a digital transformation in the industrial area. >> Muhammad: Yeah. >> This is the big thing. What's your relationship with AWS >> Muhammad: So, as I mentioned, the first challenge, data, like, we have so much data, so much computational power and it's not something that is always needed. You need it like on demand. And this is where like a hyperscale or cloud provider, like AWS, can be the key to achieve, like, the higher, the acceleration that we are providing to our customers using our technology built on top of AWS services. We did a breakout session, this during re:MARS, where we demonstrated a couple of small tools that we have developed out of our offering. One of them was ability to stream data from the vehicle that is collecting data worldwide. So during the day when we did it from Vegas, driving on the strip, as well as from Germany, and while we are while this data is uploaded, it's at the same time real time anonymized to make sure it you're privacy aligned with the, the data privacy >> Of course. Yeah. That's hard to do right there. >> Yeah. And so the faces are blurred. The licenses are blurred. We also, then at the same time can run object detection. So we have real time monitoring of what our feed is doing worldwide. And... >> John: Do you, just curious, do you do that blurring? Is that part of a managed service, you call an API or is that built into the go? >> Muhammad: So from like part of our DSV, we have many different service offerings, so data production, data test strategy orchestration. So part of data production is worldwide data collection. And we can then also offer data management services, which include then anonymization data, quality check. >> John: And that's service you provide. >> Yeah. >> To the customer. Okay. Got it. Okay. >> So of course, like, in collaboration with the customer, so our like, platform is very modular. Microservices based the idea being if the customer already has a good ML model for anonymization, we can plug it into our platform, running on AWS. If they want to use it, we can develop one or we can use one of our existing ones or something off the shelf or like any other supplier can provide one as well. And we all integrate. >> So you are, you're tight with Amazon web services in terms of your cloud, your service. It's a cloud. >> Yeah. >> It's so Capgemini Super Cloud, basically. >> Exactly. >> Okay. So this we call we call it Super Cloud, we made that a thing and re:Invent Charles Fitzgerald would disagree but we will debate him. It's a Super Cloud, but okay. You got your Super Cloud. What's the coolest thing that you think you're doing right now that people should pay attention to. >> I mean, the cool thing that we are currently working on, so from the keynote today, we talked about also synthetic data for validation. >> John: Now That was phenomenal. So that was phenomenal. >> We are working on digital twin creation. So we are capturing data in real world creating a virtual identity of it. And that allows you the freedom to create multiple scenarios out of it. So that's also something where we are using machine learning to determine what are the parameters you need to change between, or so, you have one scenario, such as like, the cut-in scenario and you can change. >> John: So what scenario? >> A cut-in scenario. So someone is cutting in front of you or overtake scenario. And so, I mean, in real world, someone will do it in probably a nicer way, but of course, in, it is possible, at some point. >> Cognition to the cars. >> Yeah. >> It comes up as a vehicle. >> I mean, at some point some might, someone would be very aggressive with it. We might not record it. >> You might be able to predict too. I mean, the predictions, you could say this guy's weaving, he's a potential candidate. >> It it is possible. Yes. But I mean, but to, >> That's a future scenario. >> Ensure that we are testing these scenarios, we can translate a real world scenario into a digital world, change the parameters. So the distance between those two is different and use ML. So machine learning to change these parameters. So this is exciting. And the other thing we are... >> That is pretty cool. I will admit that's very cool. >> Yeah. Yeah. The other thing we like are trying to do is reduce the cost for the customer in the end. So we are collecting petabytes of data. Every time they make updates to the software, they have to re-simulate it or replay this data, so that they can- >> Petabytes? >> Petabytes of data. And, and physically sometimes on a physical hardware in loop device. And then this >> That's called a really heavy edge. You got to move, you don't want to be moving that around the Amazon cloud. >> Yeah. That that's, that's the challenge. And once we have replayed this or re-simulated it. we still have to calculate the KPIs out of it. And what we are trying to do is optimize this test orchestration, so that we are minimizing the REAP simulation. So you don't want the data to be going to the edge, >> Yeah. >> Unnecessarily. And once we get this data back to optimize the way we are doing the calculation, so you're not calculating- >> There's a huge data, integrity management. >> Muhammad: Yeah. >> New kind of thing going on here, it's kind of is it new or is it? >> Muhammad: I mean, it's- >> Sounds new to me. >> The scale is new, so- >> Okay, got it. >> The management of the data, having the whole traceability, that has been in automotive. So also Capgemini involved in aerospace. So in aerospace. >> Yeah. >> Having this kind of high, this validation be very strictly monitored is norm, but now we have to think about how to do it on this large scale. And that's why, like, I think that's the biggest challenge and hopefully what we are trying to, yeah, solve with our DSV offering. >> All right, Muhammad, thanks for coming on theCUBE. I really appreciate it. Great way to close out re:MARS, our last interview our the show. Thanks for coming on. Appreciate your time. >> I mean like just one last comment, like, so I think in automotive, like, so part of the automation the future is quite exciting, and I think that's where like- >> John: Yeah. >> It's, we have to be hopeful that like- >> John: Well, the show is all about hope. I mean, you had, you had space, moon habitat, you had climate change, potential solutions. You have new functionality that we've been waiting for. And, you know, I've watch every episode of Star Trek and SkyNet and kind of SkyNet going on air. >> The robots. >> Robots running cubes, robot cubes host someday. >> Yeah. >> You never know. Yeah. Thanks for coming on. Appreciate it. >> Thank you. Okay. That's theCUBE here. Wrapping up re:MARS. I'm John Furrier You're watching theCUBE, stay with us for the next event. Next time. Thanks for watching. (upbeat music)