EXCLUSIVE: EmbedUR Spins Off ModelNova to Build AI for the Edge | Front Page

[00:00:00]Now, eight days ago, Embed UR spun off model Nova as an independent company.

[00:00:04]Nine semiconductor partners, fabulous, an open-source model zoo, a commercial platform called Fusion Studio, production ready AI models for facial recognition and wireless sensing. Now, Rajes Shubraam's own question at the launch, and here I quote, why is it still so hard to turn a working AI model into a real product at the edge? End quote. Now that question took about 18 months to build an answer to words and the answer is now its own company embed was founded in 2004 no external investment which is quite commendable 35% average annual growth for two decades Rajes you've done wonderfully well Rajes Sri Subramanyam is the founder and CEO and he's joining me exclusively on front page to well answer that question which I just read out to you and a lot more. Welcome to front page. How are you doing Rajes?

[00:01:02]>> I'm doing great. Thank you so much Sudhi for having me on this uh wonderful morning your time and evening our time.

[00:01:10]And you know it's it's interesting that the way you led it was pretty great. 20 years ago I started embed your we've been doing embedded software for 20 years. You know when when I started the company Wi-Fi was nent. There was no Wi-Fi in 2003 and 2004, >> right?

[00:01:27]>> There was no Bluetooth in 2003 and 2004.

[00:01:29]There was no Ziggby. It all came together and we just caught that wave and built a software company that specializes in embedded software and connectivity. And as we grew around four or five years ago, we realized that oh my god, was it an accident or was it fortitude or was it us preparing for compute to become commoditized and available at the edge and it was like a aha moment where it's like, oh my god, we got so much embedded talent capacity and we've built products with connectivity and now take all this. Oh, bring intelligence to it. Oh my goodness, make it run on embedded systems. And that's when we started uh tinkling and and looking at what we need to do. And what started as a simple model zoo for small models and for small devices then slowly evolved into model Nova. And like you said 18 months ago, what we started as a division has now become its own company.

[00:02:36]>> Fabulous. and serve that dedicated purpose of bringing intelligence to the core edge and I I talk about edge memory constraint and resource constraint edge environments. Oh yeah, that's such a wonderful little recap Rajes in terms of you know the entire journey that you've had and that aha moment would if I were to kind of interpret it as probably a miracle I'd just like to say this that you know when we say that miracles happen we tend to kind of correlate them to luck but it's actually not it's something which is repeatable and doing everything every day even if you don't like doing it and you've been doing it since 2004 and that to bootstrap So, well, good on you. But of course, I'd like to start here, >> you know, thank you. See, and and and and like you said, what that does is it gives you an exposure to customers, >> right?

[00:03:32]>> And when you build that relationship with customers, companies like ours get a seat at the customer's table.

[00:03:39]>> And our customers are really big customers. They are fortune 500 companies.

[00:03:43]>> Fabulous.

[00:03:44]>> Which means they have lot of R&D. They have a lot of R&D budgets. They look 10 years ahead >> and we get to sit with them and look at the same lens and that gives us an idea of where we should pivot and grow and go. Would that be probably before I get into model Nova, would that be something that you'd probably advise a lot of let's say the founders and the startups who are coming into let's say not just this ecosystem generally speaking because I mean that's probably what was your uh mindset if at all correct me if I'm wrong but when you started embed in 2004 you were very very mindful of what the customer was saying >> yeah We great great point you know so so let me so I I've learned a few things along the way through this journey and one thing was we started and we said what what we define for ourselves is what is it that we do not want >> fair enough >> let's first identify we don't want to be another software company >> we don't want to build applications we don't want to look at finance none of that we want to target connectivity we want to target Wi-Fi and we want to target embedded systems. So we stayed very truthful to what we wanted to do and we did not indulge in what we did not want to do. But then as things evolved if you think about it when we were working in 2003 and 2004 one of our customers came and said hey we want to build this CAD model and AWS had just launched >> and we want to build this and run it on AWS.

[00:05:24]>> Now connectivity was not great back then. So for us to run anything on AWS and build a small model or uh or software >> you had to have strong connectivity and so it was not possible back then so connectivity became very predominant and commoditized around 2007208 fiber optics Wi-Fi began growing all of a sudden infrastructure became the king >> which means we had to build adjacencies now we need to understand how to manage devices from the cloud so if you think about it we started started with connectivity embedded software on devices >> and then once we started moving towards cloud the cloud came in and became a management center for managing these devices which means as a company we had to start building adjacencies that will take us to that next level and say okay now we know how to build these devices how do we manage these devices how do we secure these devices so that's the product adjacency that we had to build right so >> to your question to founders I always I've learned and what I've been taught is look for your product and project and adjacencies and customer requests that will give you vertical capabilities that you can now build on top of.

[00:06:39]>> So as a result we built a cloud team.

[00:06:41]>> We built a cloud team and we started man we now managed one of the largest service providers Wi-Fi 5 network in India. It's our Wi-Fi software running there close to 8 million homes.

[00:06:51]>> Right. So, so for that you need a huge cloud capability to be able to manage so many devices, >> right?

[00:06:58]>> So, across Southeast Asia and across South America. So, that was the second adjacency >> and then the third adjacy became IoT and then finally fourth became AI. So there's been like this way where you have to capitalize on these adjacencies and build capabilities otherwise this is this whole Hollywood and uh um Hollywood there's a there used to be a store called Hollywood where you rent DVDs and blockbuster >> they never built the agency for Netflix right we never thought about cloud >> true >> and they lost their way >> right >> they lost their way so >> and they were huge companies to be honest Rajes Blockbuster was a huge company >> huge company disapp market leaders and everybody's looking at, you know, everything is running on the cloud and these companies are like >> my kids probably don't even remember about any of them.

[00:08:00]>> That's true. But yeah, it's that's a wonderful incisive point that you made that one needs to be fairly nimble and flexible without actually uh losing your core focus. But then that core focus tends to kind of branch out into other things and you need to be very cognizant about building upon those adjacencies as you just put it across which of course then brings me to model Nova now uh open source model zoo fusion studio and beta production ready licensed models nine semiconductor partners including ARM ST micro infinion NXP quite impressive by [clears throat] the way Rajes >> but here's the first question on Modernover was built inside embed over the last 18 months. Now what specifically do you feel does it need as an independent company that well it could not get as you put it across a division.

[00:08:58]Great question. If you if if we think about what we have learned in 20 years is embedded as such as we have a very core service or customer mindset.

[00:09:11]>> Mhm.

[00:09:12]>> And and and and to have a product mindset requires a separate incarnation.

[00:09:20]>> Sure.

[00:09:20]>> Because we have to approach it where we build once, sell many times. Embed your build once, sell once. Right. And then we learned that over the course of building this cloud technology that we sold as a software license agreement to our customers and we said we really didn't do a good job because we're a services company and we didn't really know how to really build a software packaged product company. M >> so as a result when we learned from that and when we saw model no all of a sudden the customer requests are coming in we are not able to handle it and and we going like okay should we treat this as a service customer a product customer and then the differentiation became very clear no this is a product it has to be licensed we have to build a structure around it and we have to build an organization around it and scale that organization very differently and that was the reason for spinning out model >> fair enough which then brings me to this aspect as far as you know what Mr. Michael Hurstston, the CEO of Lmentum and Embed UR Advisory Board member, and I'll quote, "The gap between a working model and a shipping and and shipping a product is what's holding the industry back." End quote.

[00:10:36]Please Rajes, walk us all through in terms of what specifically is the challenge between let's say a working AI model and a ship product which is at the edge.

[00:10:51]Great question.

[00:10:53]Now, if you think about models running on bigger compute, >> I'm talking about Nvidia GPUs, um, and I'm talking about, um, where memory is not an issue, resource constraints are not an issue, and it's not time critical, >> then it's okay to have those models have a level of accuracy that's not 95% or 93%. Number one, >> um, building a model at the edge is a very different ballgame >> because number one, your compute is there, >> but more than your compute, your memory requirements are very small. M >> you don't have such big memory to play with >> and you don't have lot of power >> which means you need to be able to take regular models optimize them >> and they shouldn't just work they should perform reliably >> they should be secure and there has to be a way to manage the life cycle of model of that model sitting on the device >> and the model has to be highly accurate as well. So imagine you have a door lock camera >> and the door lock camera is now running an AJI model >> which means the door lock camera running an AJI model without connecting to the cloud to recognize faces it doesn't have cloud connectivity it's all resident on the local device.

[00:12:17]>> Yeah. Now, for that model, if it's running on a door lock camera, it better be really, really accurate and work 99% of the times because if there are 10 people in the house, it better make sure that it lets all 10 people in all the time, every time, whether it's night, day, morning, or evening or dawn or dusk, lighting matters, type of face matters, >> and has nothing to do with the semantics in the family.

[00:12:43]>> Correct. Exactly. And nothing to do with Yeah. and and [clears throat] it has to operate at that resource constraint environment and it better be super accurate and better be working 99.99% of the time. So that is the production model that Michael was referring to.

[00:13:02]That's the production model that our customers want and that is that is the biggest difference >> in in in terms of being able to >> execute very efficiently [clears throat] with very low latency and make decisions quickly and be right every time. But then tell me Rajes I mean since you just mentioned something like this and it's a wonderful example that you just gave a door lock example for example a door lock camera example now that needs to run probably for its own entire life cycle let's say about I don't know 5 years and and and and the kind of times that it needs to be accurate for is 99.99%.

[00:13:38]My god, those are huge challenges in terms of ensuring that you're getting it right bang on every time, time and again because let's be honest, that is a very case sensitive sort of a situation in everyday's uh in in every person's everyday ritual sort of an uh situation.

[00:13:55]You cannot you cannot irate the customer there because then you lose credibility >> and you lose it immediately >> and then the customer coming back to you almost becomes next to impossible. What have been the kind of challenges Rajes that you've faced in the past and and what are the learnings and because let's be honest challenges are nice they answer a lot of questions but then to get it bang on tickity boo every time in a production grade situation is a whole new different animal please walk us through that see when when artificial intelligence moved down to the smallest lower low power devices the idea was that we can split compute and not have to rely on the cloud all the time.

[00:14:43]>> Yeah.

[00:14:44]>> And also make it more secure because it's more there's more privacy because nothing is shared and nothing can be hacked in from the cloud.

[00:14:53]>> Fair.

[00:14:53]>> And and and there is also lot of um the ability to also scale it to different platforms >> as and when needed. Uh because if I'm a manufacturer, I'll give you the other other side of it as well. Because if I'm putting out door cameras today, two years from now, I'm going to get chipsets that are like half the price, which means I have to pass down the cost savings because otherwise my competition will >> which means I need to make sure my ecosystem is important. So I can take the same thing that I built here in terms of models, in terms of ecosystem, software and everything and quickly scale it and move it to another product which is my next generation, my next product that goes out into the market.

[00:15:36]>> So as a manufacturer I need to be cognizant of the fact that I've built the right framework and also understand hey is this that's again I'm talking about our customers have some genuine questions right when we talk to them they say hey it's great we're building it for this next 27 and 28 But then I need this working on this next two generations of platforms in 2030 and 2031.

[00:15:57]>> Right.

[00:15:58]>> Right. So which means now hardware is always 18 months ahead because hardware comes out first >> then software comes out. And now when software comes out you need to make sure that with AI the biggest thing is validation >> right >> and the the the the maximum amount of time taken is in creating the model is one thing but validating the model works and then finding out the conduct cases where the model doesn't work >> go back and retrain the model >> and then get to reroo the whole thing and now it's production ready it gets deployed the model continues to learn and then this model needs to be then upgraded over its life cycle >> and then continuously needs to be refreshed. So all these are real world problem. Oh not only that oh the model needs to be secure >> most important >> we don't want most important the model we don't want any other form or shape or way for this model to be corrupted >> or to be hacked into or to be updated.

[00:17:05]So which means now there is also the software management and this goes about it's not only the model that you own you own the software you own the ecosystem you own the application you own the maintenance you own the life cycle so these are the bigger challenges customers are already thinking oh my god we never had to deal with it at a microcontrolling right I mean nobody even thought of models running on MCUs now this is a this is all new for everybody it's happened in the last 28 36 months >> and products are slowly getting rolled out and it's it's it's it's a journey that all of us are going through and these are the pain points that we're working with and trying to solve for our customers. So it's a constant work in prog progress. That's how the way I understand it. And of course you know then this comes to mind Rajes I don't know if you've read up read up on this very recent uh it was quite an unfortunate incident which came about where you know there were a certain amount of CCTV cameras which I as you are probably obviously a institution on the fact that you know edge AI devices that entire China conversation etc and things like that but in terms of India going ahead and utilizing those and then there was a almost like a blanket ban on most of everything but then there was another very unnerving incident that took place. Apparently there were a couple of CCTV cameras which were uh in an audit uh final uh they were caught in terms of across the border. They had found a little space in that chipset where they could hack into and they were getting live feeds across the border. So you know that just made me wonder about the fact that you just mentioned secure.

[00:18:41]I'm thinking that would be at the top of your mind every time right before you start to build or you do anything.

[00:18:49]>> So 100% and majority of these issues come from the fact that it's connected back to the internet and everything is >> yes >> connected back to the cloud and we all I mean let me ask you a simple question because we deal with Wi-Fi and connectivity in the house a lot >> correct. What do you think is the average number of clients that connect to your Wi-Fi router in your home?

[00:19:12]Take a guess.

[00:19:14]>> I don't know. Maybe about one. Sorry, about 10, 15, I don't know.

[00:19:19]I'm just stabbing at it.

[00:19:22]>> Great. Great guess. India used to be four, but now it's slowly growing up to 10.

[00:19:29]>> Wow. The US The US is anywhere between 25 to 45.

[00:19:37]>> Wow.

[00:19:39]>> And the household doesn't even know that some of the clients are connecting to Wi-Fi.

[00:19:46]>> And there in lies the problem. You don't have control over all these client. It's great that we have connectivity.

[00:19:52]>> Agree.

[00:19:53]>> It's wonderful. We have Wi-Fi in every one of these devices. Everything that you buy today has Wi-Fi.

[00:19:58]>> Yeah. and you just go in there and blindly configure it, set the default password and let it connect to your internet router >> and then it voila, it goes to the cloud.

[00:20:07]>> So 90% of the times users don't change the default password on these devices, >> which is true, >> which means admin is the username, admin 1 2 3 is the password, >> right?

[00:20:19]>> It doesn't take a rocket scientist to hack into a security camera if I get an IP address. I can do it today. I can show you 65,000 cameras across the world that I can actually hack into and actually look at the feed live because there are so many people who have not changed their username and passwords.

[00:20:33]>> Wow. So that's basic hygiene which is missing. Is it because of ignorance?

[00:20:38]>> It's it's it's because we feel there is no threat.

[00:20:43]>> We wait for the bad things to happen >> till somebody acts into our camera. I mean and and the chances of that happening to a normal human being in a residential household is probably 2% or 3%. But imagine a security camera in a bank >> and if I can target a bank right and financial institutions or something better. Now that that's again going back to that that is the fundamental problem and it needs a lot of customer education.

[00:21:08]>> It's great to adopt technology but you also need to understand the evil that comes with it if you don't manage it well.

[00:21:15]>> True.

[00:21:15]>> Number one. Number two, now as we go and look at it in the next 3 years, you're going to find cameras that you can buy and stick to shelves or stick on on your on your door >> and it can monitor the footfall without having to connect to the internet and just connect through Bluetooth and let your phone know what's happening. No internet connectivity needed, >> right?

[00:21:38]>> So things are going to change and we've learned a lot. See as as as you guys as as well as manufacturers, everybody is learning a lot about the ecosystem and also compute is getting cheaper.

[00:21:52]>> Correct.

[00:21:52]>> I mean if you think about it your iPhone the the processing power that you have in your iPhone used to be what the Mac had like 8 years ago >> or 7 years ago.

[00:22:01]>> True.

[00:22:01]>> A MacBook. M although Apple is trying to escalate the prices because of the Apple intelligence bit because of again the compute capacity situation which Steve Jobs said that he would probably be you know reaching out to or or or extending that to the customers.

[00:22:18]>> Yeah, absolutely. And now we also have this whole memory issue because memory is being consumed by data centers. So all the other people are not. So so there's the whole slew of problems that are happening. But the the point I was trying to make was compute is becoming available at much lower cost >> today. So like a microcontroller or a variable as compute that we probably didn't even think would we were using that kind of comput in a laptop 10 years ago is now available on a variable in a in a small edge device. So when you have that level of compute and when you have intelligence combined with it, why do you need to push everything to the cloud >> and you can make decisions at the edge, >> it brings up a huge shield of security already. It then helps you make sure that nothing goes to the internet.

[00:23:10]Nobody can hack in through your Wi-Fi router or internet and get into the device. Not possible.

[00:23:14]>> The only way that is going to happen is if it goes through your phone or Bluetooth. M so essentially Rajes then uh moving forward would that be for example your winning pitch to anyone and everyone as far as enterprises etc is concerned uh so to say will that be the edge in a way to kind of you know promote what your overall thought processes as far as the vision is concerned which is trust and security.

[00:23:44]>> Yes. So but with a caveat the caveat being when you look at IoT industrial variables >> that is where modern Nova is going to play. We not talking about see if you think of edge there is Intel is also at the edge Nvidia is also at the edge edge compute >> and and you look at AMD Broadcom they all all at the edge. I'm talking about processes that cater to IoT variables, industrial agriculture, things like that where you know it's microcontrollers with artificial intelligence and a AI accelerators. We are now focused on that segment.

[00:24:21]>> Now that segment doesn't need all that connectivity is what I'm saying. That segment can run models at the edge itself. Which means whatever inferencing was happening in the cloud is not necessary anymore. inferencing can happen locally. Latencies are much much quicker, the lower latencies, decisions made quickly and it can speed up industrialization. It can speed up variables. Think of a simple um you know, have you heard of the Sonic toothbrush?

[00:24:50]>> Yeah. Yeah.

[00:24:50]>> Um it's got Bluetooth today, right? It's like $29 $290, $300 and then you can use that toothbrush and and I have I have one of those and it's got a Bluetooth.

[00:25:00]It's got AI model running on it and it actually will tell you how you're brushing your teeth and then onive at all.

[00:25:09]>> It's not it's it's like regular toothbrush, >> right?

[00:25:12]>> And and it doesn't it doesn't actually it doesn't it's not invasive. It tells you that you're like brushing too hard or brushing too soft and then as you turn on the AI, it learns your brushing model and then it gives you a map of where you've not brushed properly. And imagine if if that can come down to $29 and $25 and it gets commoditized and everybody can buy it. It creates a huge impact.

[00:25:36]>> Oh yes.

[00:25:37]>> Because then you democratize technology and computer is going to get there.

[00:25:41]That's that's my point. That's what I'm saying. The next 5 years >> it's going to be affordable. It's going to be a game changer >> and these models are going to run more efficiently. So it's it's not the only person who's got $290 who can afford it, but somebody who's got $20 can also afford it. And that is the kind of edge technology that's going to make a difference and that is the pace at which we are seeing technology move.

[00:26:06]>> Fair point and of course I have to say this Rajes it'll make a huge dent as far as dentists are concerned. I had to say that. [laughter] >> No in fact you don't have to go through a root canal right treated before it happens.

[00:26:21]>> You don't have to go down that route at all.

[00:26:24]Yeah, you don't have to go down the route at all. I mean, it it's could be it could be good news for dentists because I'm sure even they don't want to go through those elaborate procedures and and create pain for their patients and it also gives them an opportunity to say, hey, do the do things better, better hygiene for all of us and maybe more people coming to them. Think of it that way.

[00:26:43]>> Makes sense. Makes sense.

[00:26:44]>> And just tell them, am I doing the right thing? Yeah.

[00:26:46]>> Makes sense. That is the whole truth.

[00:26:51]>> That is that's true. Yeah.

[00:26:53]When we talk about compute, you and I mentioned it as as you did Rajes. U again I'd like to bring your attention to the fact that you know let's say a Google is investing $15 billion in Visag. Now Nvidia GPU superclusters. Y is deploying 20,736 Blackwell GPUs and embed UR is building AI that runs on a chip without cloud connectivity. That's where you're headed. Now I would request you to please let's say guide us all to understand where AGI fits. Is it replacing the cloud competing with it or is the $15 billion investment in Visa actually what makes model Nova possible?

[00:27:37]>> Great question. You know see at the end of the day we have our own uh small data center. Model Nova has its own small data center where we have GPUs ourselves.

[00:27:48]Why do we need that? Because we need to first take a model and train it on these bigger GPUs >> and then make sure the model works in the bigger scheme of things. So once it works in the bigger scheme of things, then we quantize the model. So even for us, we still need higher compute.

[00:28:08]>> Fair enough.

[00:28:09]>> To then quantize compute to get to the smaller devices. Now once we get there, we still need the ability to somehow upgrade these devices, manage these devices and look at what software they have. And this is the life cycle management. So the life cycle management still needs the cloud.

[00:28:30]>> Sure.

[00:28:31]>> What it doesn't need is the constant back and forth where >> oh capture, send, inference, send back decision. M >> so this is a constant cycle that's happening today.

[00:28:46]>> Now what's going to happen? What is going to change is capture inference decision done. Oh >> inference decision done anomaly sent >> that's the time we still need >> right.

[00:29:04]>> Oh then I've learned learned learned I've sent you these new things that I've learned. Oh, the model gets updated back in the cloud and now we need to find a way from the cloud to push the model back. So cloud can never go away.

[00:29:17]>> Got it. But the dependency on it over a period of time goes down.

[00:29:22]>> The frequency of >> frequency. Yes.

[00:29:24]>> You know communication frequency is going to rapidly reduce which means the power requirements also going to rap because every time you have to transmit up and transmit down you would consume more power.

[00:29:34]>> Correct. And so that is so they have to coexist except it's going to be lesser dependency in terms of influencing decisions because it can be made locally.

[00:29:45]>> Makes sense. So essentially what are your thoughts uh Rajes as far as the entire compute capacity thought is concerned. I know you've spoken about the fact that you know compute generally speaking has become a lot less in terms of the commercial factors but compute capacity is genuinely a serious constraint and how do you see it impacting uh the overall road map ahead because let's be honest the AI demand is only going to keep increasing manifold as we keep moving forward.

[00:30:18]>> Yeah. See that is I I see I think that's a dualprong problem right because on one one end you need massive infrastructure to install that compute it's just not being able to get access to your GPUs it's also getting access to power access to cooling access to infrastructure access to uninterrupted infrastructure so that is something that we are all trying to figure out because this all happened it just it just came upon us like >> in the last two years when I first got three years when we heard about chat GP is like what just happened >> right and so now we're all learning together to see okay how can we manage the compute that we have how can we scale so it is got to be a industrydriven model and we we're all still learning and I think it's going to take another 12 to 18 months for us to figure out what is ideal for us because again there is also a lot of wastage I think we've not learned to eliminate what is not necessary.

[00:31:19]>> I was actually >> just doing everything thinking everything's >> Yeah.

[00:31:22]>> Yeah. And and and we will learn what we can eliminate. Do we really need this level of compute to do this training?

[00:31:28]>> Exactly.

[00:31:29]>> No. Let's get smarter at >> and so we'll evolve because we've evolved from that perspective. And I think that evolution is happening now and in another 12 18 months we'll all get smarter because anything that's rare >> we find a way to circumvent it and make sure that we use it like it's rare.

[00:31:48]>> Um so that is what we are also learning and and today for us also we would love unlimited compute >> and we I would I would love to launch 30 different production models to our customers but we are also limited by the comput. So what we go is look at the top three verticals. These are highest priorities. Let's work on these.

[00:32:06]Eliminate the the bottom.

[00:32:08]>> Fair enough.

[00:32:09]>> Bottom 20, right? I mean because 20% of what you make >> gives you 80% of your impact.

[00:32:17]>> Agreed.

[00:32:19]>> Agreed.

[00:32:19]>> So the 80/20 rule, the PAT rule applies always. So we go and look at okay of 10 things give me the top 20% that will give me 80% revenue impact profitability impact and customer acceptance let's go focus on that that way we decide limit our compute capacity to this and limit our capital expense to this >> yeah because both of them go hand in hand and I I don't know if you agree with me on this one uh Rajes but what you made was a very valid point that everyone is going through that uh you know learning curve right Now, but do you agree with me when I say this that then it becomes a very software ccentric sort of a a mindset moving forward where you'll really need to a lot of the big frontier labs probably or everyone will need to really work on the software aspect of things for them to reduce the load on the compute capacity like you mentioned yourself if you want to go ahead and talk to a particular model about the fact that listen I need to know how to bake let's say a carrot cake right you don't need the biggest compute capacity in answering that question. VCB there is a developer ecosystem which is trying to solve a very hard problem when it comes to coding. So then the layered approach do you agree maybe that could be the solution moving forward?

[00:33:34]>> It is you absolutely nailed it. That is the solution moving forward. Right? I mean, instead of having a large language model sitting in a big server answering questions about where you need to go in an exhibition floor, think of it sitting in a small edge device with like probably maybe 512 megabytes of RAM feed the floor space. Now, it talks to you, doesn't need cloud comput, doesn't need any inferencing, and all of a sudden it tells you where to go and it's it's got this agent AI running on it which tells you, oh, which floor, which level is this booth that I need to go to in this big exhibition center. So we are splitting compute between cloud and between these devices. As a result, we don't need to rely on all the big compute all the time.

[00:34:16]>> Fair point. Which then brings me to the aspect of the tiny ML which runs AI on let's say microcontrollers with as little as 256 kilob of memory. Now a hearing aid detecting speech for example, a smartwatch detecting arterial uh fibrillation, a soil sensor telling a farmer when to irrigate. So please again shed light on this Rajes uh in terms of the lower boundary of where AI can actually run today and what becomes possible when intelligence costs less than the chip it runs on.

[00:34:52]>> Yeah. See the the biggest the biggest thing is now let's forget compute for a moment.

[00:34:59]>> Let's forget uh let's forget the memory for a moment.

[00:35:02]>> M >> let's focus on power. H >> and if you want to run and I'll give you an example your simple continuous glucose monitoring system I'm not sure and Europe so you like >> various people use >> the battery lasts for 14 days the the the product lasts for 14 days correct >> right >> but it can last for more than 30 40 days >> the reason it lasts for 14 days is nothing to do with power but it's something else >> I'll not get into that topic The issue is software and the ecosystem and the environment around it has to be sensitive to power >> which means you need to have two modes.

[00:35:47]You're sleeping most of the time.

[00:35:50]>> Yes.

[00:35:50]>> Not consuming power. You sense wake up do your job. Go to sleep.

[00:35:57]>> Makes sense. So that way that that way you do you you deploy power conservation.

[00:36:05]That way you increase battery life. So again imagine a camera that runs on battery that you can stick on top of your door >> that can do footfall counting for you.

[00:36:16]>> Correct?

[00:36:17]>> Imagine a a fabric store in in the middle of New Delhi.

[00:36:22]>> Pretty soon they'll be able to buy these cameras or even sensors. Let's say cameras where they'll like it's got a sticker. You peel the sticker, you just stick >> and all of a sudden it can record footfall and within a month tell you which time of the day records your biggest footfall and what days they are so that you have you man the floor better on those times and those days.

[00:36:44]>> Fabulous for the business.

[00:36:47]>> Fabulous for the business. Right.

[00:36:48]>> Yeah.

[00:36:49]>> Cost probably say $3 >> right now. That camera doesn't need to be on all the time. It has to sense motion. So it has to have a motion sensor. It senses motion, wakes up and then as an outline of a person walking in, counts, sleeps. So this is the level of sophistication that these devices are going to get. They're going to get that level of software, that level of intelligence, that level of sophistication. So you have to conserve power. It starts with conserving power.

[00:37:16]>> Fascinating.

[00:37:18]>> All the Yeah. All the chip vendors are looking at this saying, "Hey, how do I give you more capacity over a longer period of time?"

[00:37:26]>> Imagine you're talking to one of our customers and this customer has got a label, a sticker >> that you can stick. It's got Bluetooth in it. You can just stick like a regular label on on your package >> and it has life and battery power for 1 and a half years. [snorts] >> Brilliant. and and we're working with them to build a tracking solution and build a model around it >> so you can track these packages seamlessly >> which means that label >> has to work for one and a half years and it's it's going to work. So you can put it on pellets. That pellet has the same label for one and a half years. Now you can track the pellet wherever it goes.

[00:38:04]You know you got you have software then it beacons off. You can look at how the pellet moves in moves out. You have now a model on how often a warehouse brings pellets in takes pellets out. Think of think of another example. We have this wireless sensing model that I talked to you about, right?

[00:38:21]>> Uh where we can sense motion and one of one of the one of our customers is looking at an application where they have these frozen containers that they take take in these big uh transport vehicles >> and sometimes the door opens. Now they don't know if the door is open because it's manually has to be bolted shut by somebody right >> now. If it opens and the guy is driving, he never he never knows that it's open till he gets to the end. Now it's frozen product. Now it's damaged goods by the time it reaches. Imagine he's got a dashboard while he's driving where the motion sensing model has just sensed the door is open. He can immediately stop, go make sure the latch is on. And imagine if this costs like a few dollars to implement.

[00:39:05]>> That's fabulous Rajes. I mean those are real life solutions you're talking about. I'd like to I'd like to follow up with something uh which which you just spoke about which I thought honestly is brilliant in terms of the fact that power is which is at the core of this entire understanding but generally speaking the industry mindset and I'm not speaking about you I'm saying generally the large enterprises etc whatever you know the general mindset is wait man why should I give the consumer something uh which will last them let's say for about 2 years that will hit my sales it'll probably hit my volumes rooms. So then, you know, let me kind of make the life a little shorter so that you know, I can always keep them hooked on and my production levels are keep going up, etc. and things like that. Do you think that might be a challenge when you're talking about these things moving forward? Because I think that could be one of the reasons for that 14-day period.

[00:39:56]>> I don't know.

[00:39:57]>> Uh uh the so the reason for the 14-day period is more biological because the sensors get corroded by the body >> and as a result, you have to throw it away. the battery itself can last longer. But to answer your question, look at look at the flip side.

[00:40:13]>> If I can make something last longer in terms of hardware shelf life, >> Mhm.

[00:40:19]>> what is the value addition that I can build on top of it and monetize it more on the software side?

[00:40:25]>> Makes sense. So all of a sudden I get you to now buy services from me as software >> because I'll tell you how soon it's going to die or how long it's going to last.

[00:40:38]>> Now you would rather buy insurance.

[00:40:40]Human minds we always want to buy insurance hoping the worst will or fearing the worst will happen >> and hoping for the best. Now if if we can monetize the mindset of I will give you a lot of software value addition you just buy this once it's going to last forever and we will give you all this software souped up stack on top of it and give you all the additional services all of a sudden the customer mindset is going to be like I only have to spend this up front >> makes sense >> that's it I don't have to pay that $100 it's only five oh and I only have to pay $5 every year for the next 25 years it's an Easy sell, >> right?

[00:41:20]Makes sense. Makes sense.

[00:41:22]>> Yeah. So, because we never look at it as we look at it all of it is capital cost when we look at, oh, I have to buy this whole thing now. No. And that's where it's like, okay, no, you don't have to buy the whole thing now. It's cheap.

[00:41:32]It's like there, you know, oh, by the way, you get all this additional thing.

[00:41:36]You get notified when the sensor fails.

[00:41:38]You get notified when the sensor is included. I'll buy this. I'll buy this.

[00:41:42]All of a sudden, I'm buy spending more than what would I have done otherwise.

[00:41:47]Fabulous. Interesting. I like playing the devil's advocate once in a while.

[00:41:50]Rajes, I hope you enjoyed that.

[00:41:52]>> No, I I I know, but I I I I got to I I I got to chime in. The way we we're talking to customers and what they're telling us, I'm just passing on. I'm just a messenger. [laughter] >> Fair enough. Which then brings me to this again. I mean, I'll come back to you know the the the sovereignity aspect of uh the discussion. I mean, it has been front and center lately. Rajes, I'm obviously sure that you're aware of that. Now, processing on the device eliminates sending sensitive data to the cloud as we've discussed at length. Now, the DPDP act, mate's five layer sovereign stack, the fable mythos export control which just happened. Now, Agi is the physical answer to the data sovereignity question. So, Rajes, my point to you is this. How central is that argument to how embed you are and model nova sell today?

[00:42:48]>> It's very important right? It's one thing to have a model. It's one thing to make decisions based on that model >> and it's another thing to say who makes the decisions >> and who calls out. Okay. So we've built this edge intelligence system and let's say we've just built this intrusion detection system or a false let's let's talk about posture posture detection in senior homes right >> where with the with there's lot of models that we are running right now for posture detection. Now it's one thing where we can license the model to a senior home that will buy the software from us >> but then the actions that the model has to take has to be defined and has to be built by that ecosystem and by that application which means the sovereignity of that whole experience on actions the model needs to take is dependent on industry action and as a community and as an industry they this needs to be defined on A B C and D that needs to happen. This is possible. This is not possible. And now this needs to confirm to overall governmental rules. What works? What doesn't work? So it's a very important question. Again, we rushed into all this model development. We've rushed into all of this. We have all the cool things right now. But right now, as we talking to customers, as we're talking to applications, different applications have different kinds of problems. And so you look at applications, you look at industries and you look at countries >> especially becomes very very important.

[00:44:26]>> If India has to be sovereign and build something unique, India has to own India's models.

[00:44:36]>> I agree.

[00:44:36]>> India has to own the outcome of the models.

[00:44:39]>> India has to decide the confirmance of this models. India has to figure out what that software stack is going to be running on that model.

[00:44:47]>> So, so it's a it's a super important question when it comes to industries, verticals, cities, countries and governments.

[00:44:56]>> Rajes, you you've been somebody who has been very very uh important as far as the as far as your voice is concerned in the entire ecosystem. Uh you of course have been speaking a lot with your peers. What are your thoughts? Because I remember I mean I I'll I'll bring this to your notice. uh Vive of Sarvam uh he and I spoke on front page and I remember him speaking about something which I thought was quite interesting. He mentioned the same thing which is about owning our models and he said that it doesn't matter we shouldn't just go uh you know hammer and tongs in trying to go ahead and build the frontier model.

[00:45:30]He says how about we go with frontier minus one we go with frontier minus 2 but at least there is constant regular progress towards it and then we reach a point where we have will have our own models. I just thought I I'll bring in that little uh understanding in your uh entire ecosystem. What do what do you say to that? No, >> I I totally agree with that because you you cannot run till you know how to walk.

[00:45:56]>> Agreed.

[00:45:57]So and and and you cannot walk till till you know how to crop. So we we are still figuring out with a massive population >> Yeah.

[00:46:08]>> and a a ecosystem and um group of talented engineering talent that is outmatched.

[00:46:20]>> Yeah.

[00:46:20]>> In in terms of in terms of what the world holds today, right? I mean average age of around 28 or 29 strong engineering background and talent >> and we are all adept and and in fact the country is adept at software engineering it's been the it's it's been the center spine of what if if you look at India today right it takes take great pride in in saying that >> a lot of stuff that happens there gets shripped globally through customers all around the world >> that's true >> and and and and and and to build something unique and to build something for the nation and for the country, India needs to make sure that it does take those steps first. You don't have to build the frontier model. Build the first model that works in that ecosystem and then learn from it. build the next one and build the next one and finally get to the final step. And we will learn a lot in the process because it's not easy >> to build something for a country with such a diverse population with such a diverse language set and such diverse you know centers of excellence across the country >> and uh and and it's a very challenging problem and not many other countries face that >> you know and and So I I can build one one model that is only dependent on English and run it across the board. It works everywhere.

[00:47:49]But I cannot take that and scale it in India.

[00:47:52]>> Which is true >> and and for that reason I totally agree with them. Let's let's target let's first take those steps. We learn from the steps and then we'll scale because scaling comes naturally in India.

[00:48:06]>> It does. That's true. No other country can say we have 480 million cell phone subscribers.

[00:48:12]>> Correct.

[00:48:13]>> Not possible.

[00:48:14]>> UPI Rajes. I mean it it it up.

[00:48:17]>> Unbelievable.

[00:48:18]>> Yeah.

[00:48:19]>> Yeah. That was a baby step. It started and look at the adoption now.

[00:48:24]>> Yeah. Makes sense.

[00:48:25]>> And and and we know it. We we've done that before and and we've seen that grow before and it can happen again. We just need to be cognizant of it >> very quickly. Do you think we are in the right direction though?

[00:48:39]>> Absolutely. Absolutely.

[00:48:41]>> Wonderful to know.

[00:48:42]>> Empathically, empathetically, empathetically empathic right direction.

[00:48:47]>> Um investing in the right areas looking at bringing semiconductor to India focus only about a week or so back. I'm sure you're aware of this. Semicon 2.0 1 lakh 27,500 K rupees has been allocated and released. So yeah >> abs absolutely even there we targeting what works for India industrial automotive healthcare variables we don't have to target the two nanometers we don't have to get souped up engines goes back to the question you asked me about how do we do we go we don't have to go for the frontier 2 nanome sophisticated GPU buildup no we start here because this applies across India because that's what India needs today for industrialization and growing industrialization in terms of edge intelligence is what I meant >> and and we are focused on the right sectors.

[00:49:38]>> Fabulous Rajes. Then of course we spoke about talent. Uh Modern Nova is headquartered in Chennai. So brilliant wonderful talent there. 249 engineers in early 2024. I might get the numbers wrong so please pardon me for the same.

[00:49:54]420 plus today, 550 plus by the end of 2026. Now Model Noa independently is headquartered in Chennai. What does Tamina do? Give embed UR in model Nova that let's say California cannot.

[00:50:10]>> So, so let me rephrase that. Model Nova is headquartered in California with its main operational center in Chennai.

[00:50:19]>> Sorry, I got that wrong.

[00:50:20]>> Model Noa is still a US company.

[00:50:22]>> Okay.

[00:50:22]>> With the headquarter in US and our biggest representation is in in Chennai.

[00:50:27]>> So, I got that half right.

[00:50:29]>> You got to give me that.

[00:50:30]>> Yeah. Yeah. you you to you totally got it half right. Um and and and the and the biggest answer to the question is 21 years of being in Chennai as embedded or India >> we've built strong relationships and the state by itself produces the maximum number of engineers >> by any standard in the country >> and we we've seen that and uh maybe I'm a little partial to that too. So um and and we've kind of created a culture where we built relationships and we participated in the overall university ecosystem to model it in such a way that the talent that we hire from them is actually pre-trained by the time they get to us.

[00:51:17]>> Wonderful.

[00:51:18]>> So again this this is where we don't do everything that everybody does. There are a lot of other companies that do everything and then they hire people.

[00:51:26]there is something called uh Ben. So we we don't go through any of that process.

[00:51:32]We work with a select group of institutions. We focus on what the skills are needed for our industry. We work with them exclusively and then we bring those people into the talent pool into our organization. We train them and we train them through our US engineers and then train them through our Indian engineers and then we make sure that they start participating very quickly in enabling what we want to do in the AI space and what we found is extremely great talent and very quick learners and very open mindset eager to learn eager to scale and um great great enthusiasm.

[00:52:12]Fabulous and the fact that you know a lot of uh people who have come on to this show and I as well maybe believe in this Rajes that you know Tamminadu has this a wonderful little trinity of sorts you know the ports the talent and the ecosystem with with companies like let's say the Foxcon and Pegatron uh you know cluster which has done wonderfully well >> you know the entire automotive ecosystem etc. So do you do you agree with me on that?

[00:52:41]>> I totally agree with you. See, I I personally feel that there are a lot of other states that have the same potential and we seeing that happening now, right?

[00:52:50]>> We seeing the gift city concept that we looking at in Gujarat, right? We we we love that concept. It's it's a pretty unique concept. We want to do something there as well.

[00:52:58]>> Bangalore Bangalore is hub of talent.

[00:53:00]But the problem is Bangalore has got so many so many great companies that attrition is a big a big issue right so it's it's not that it's it's not that we've not experimented but this is kind of relationship that's been built and we've been able to manage it here but we feel that gradually the rest of India is also picking up steam in these areas >> is nice to know and the fact that you know when you have an IIT madras that kind of in any case puts you ahead of race because my god the kind of work IoIT Madras does as far as deep tech is concerned is truly commendable >> pretty it's pretty it's pretty it's pretty phenomenal it's pretty phenomenal and we've been blessed to be associated with some of those programs as well >> fabulous finally before I let you go Rajes I have to ask you this AJI software market $2.4 billion in 2025 projected to be $ 8.89 $89 billion by 2031. Aji hardware market $33.3 billion in 2026 projected to be $81.12 billion by 2032. Now two companies uh where there was one what does embed look like in let's say 5 years and what does model nova look like if it gets this right which I'm pretty sure it will.

[00:54:18]model Nova based on the traction that we seeing in the amount of customers that we are signing up I think we are going to beat our embed your growth profile of 20 years in just in the next two to three years >> congratulations on that well done >> yeah thank you it's it's just crazy I think sometimes you get ahead of the curve sometimes you catch it at the right time uh we've been doing this for 21 years and we've always been following the leaders and for a change this time we thought we lead and it's working out for us >> fabulous Before I let you go, finally, edge AI.

[00:54:52]If you were to define it in a line, how would you do it?

[00:54:58]Edge AI in a line.

[00:55:02]Game changer.

[00:55:04]>> Think no. Um, I'm going to say I I just thought of it. Think of everything that you touch and use. Imagine if everything that you touch and use as a brain, how will it behave?

[00:55:20]>> Well said sir. Well, there you have it ladies and gentlemen. Mr. Rajes Subramanyam who started off in 2004 had a wonderful vision stuck to his track bootstrapped kept pivoting kept building adjacencies was nimble enough and I believe through this conversation you all would have gathered the fact that he still is nimble and most importantly is hungry. Please do let us know what are your thoughts in the comments below.

[00:55:50]This ladies and gentlemen is front page by the IM network. Like, share, subscribe and always remember thinky I think I am.