The Future of AI: At the Edge with Latent AI’s Jags Kandasamy

This is Textron tv. Hey everyone. Welcome back here to another Textron TV segment.

Our guest for this segment is the co-founder and CEO of latent ai. His name is Jags Kasami. Uh, Jags.

Welcome to Drug tv. It's great to have you on here. Thanks, Ellen.

Thanks for having me. So exciting, Ben. Fantastic.

So, Jags, I mentioned you were the co-founder and CEO over at Leighton ai, but you know, there was life before co-founding, right? Leighton, let's hear a little bit about your life pre latent ai. Absolutely.

Um, right before, uh, founding latent ai, I was running, uh, another startup called AutoSense, where we had automated human hearing. Imagine this, right? Listening to sound and then identifying events from it.

E uh, took the company towards, uh, manufacturing and the predictive maintenance applications. So li you know, if you know a good mechanic before they touch a machine, they listen to it, then figure out what's wrong, and then they, they approach it, right? So that's basically what we had automated.

So we were in, uh, in, uh, uh, working with Delta Airlines, with, uh, Ford Motor Company with Airbus and several power plans in, in Japan, right? So being sound to identify things. One of the, the issues that we, we had, you know, this was the first Edge AI component, right?

In the mid 2010, uh, everybody was thinking about ai and we had taken AI to the edge, right? Running it on the, uh, on, on near devices, which makes sound. And that's where, uh, uh, you know, I got my start.

And prior to that, I was at H-P-H-P-E, started as an engineering manager, and then ran customer success, customer support, and did several things. Uh, at HPE, my first company I started when I was 21 back in India. So really been here for 20, 25 years now.

And, uh, you know, that's my journey Been of a lifelong entrepreneur. Very cool. So interesting.

With the company, with the sound, it was sort of machine learning where you matched patterns of this sound sounds like this. And when it sounds like this, this is what the, the situation usually is, or this is what Absolutely. You know, the status is.

Yeah, we, we actually, uh, took every chunk of sound you could. Uh, the chunk could be one millisecond or five seconds. However, however, uh, granular you could go, uh, you wanted, you can do that.

And then we, uh, went around, uh, analyzing each chunk in three domains. Frequency, domain, domain, amplitude, domain and time domain. And then, and we extracted some features.

That was our core algorithm, our, our proprietary algorithm, and we extracted features, and then we did the pattern branching on those features. So you got similar sounds to group together and we were able to easily visualize and, and, and, uh, tag them and, and push them out into the model. So, Excellent.

So you went from that to founding latent ai. What was, you know, as I, I spoke to you off camera, no one founds a company lightly, right? We all put our heart and soul into it, blood, sweat, and tears as well.

What was driving you that latent AI was going to be a, a must have a game changer? What, you know, what was, where'd that passion here come from? So, a couple of things, right?

I have to go back to the Artan story. When we started AutoSense, um, imagine collecting sound, you know, we put a sensor out there, collected sound, and then, you know, you started collecting everything in the cloud. We were using a hotspot early in the days when you use a hotspot, there is a bandwidth limitation.

We got a notification from our telco provider that after two days of collection, your bandwidth for the month is done. So that kind of brings like, okay, you need to move things to the edge to, to analyze. That's one.

The second issue that we hit was, uh, the company was ultimately acquired by analog devices. When analog devices came to us and they, they invested in us. One of the, uh, uh, the memorandum for us was to run the algorithm, which we were running on an arm processor to be run on their blackfin DSP, their digital signal processor.

We literally had to rewrite every line of code from arm to TSP took us seven months. Once we finished that, a DI said like, alright, we are acquiring it. Right?

They wanted to make sure that it ramp That it would work. Yeah. So after the exit, I, you know, the things that was, uh, bothering me was that one, why did it take so long?

Why do we not have a universal compiler to bring AI models to different hardware devices? You know, uh, in GCC compiler, we, we basically switched the flag and we are able to compile quickly, right? For, for regular application, why can't we do that on the ML side as well?

That was one problem. The second one was we kept increasing our bandwidth from, you know, 3G to 4G to 5G from a, from a telco perspective. And then we went from, you know, a coaxial cable with, and the broadband to, to fiber and stuff.

But still the applications and the demand kept going increasing, right? We, we are never happy with what we get, right? The, the, the theory of induced demand.

So those things came into my mind, like, if you want to process these things, your processing needs to happen at the source of data. Whenever and wherever analog to digital conversion happens, that is where processing needs to happen. That is where AI needs to run.

That was a, a a, a nirvana moment for me in the, in the mid 20 fifteens, right? And I wanted to, uh, ensure that we kept that and I, I was solving for that problem in a continuous manner. And that was what the origin of leading AI was.

When I joined SRI International and Stanford Research Institute is the, uh, birthplace of c nuance communication. And it is a place that internet terminated. Uh, back in the seventies, right?

When, when, when our darpanet, uh, started out, I I I, I went to SRI as an entrepreneur in residence and met with my co-founder, s Zach had worked on technology that compressed neural networks, right? So, okay, the compression is required to push things to, to the edge. And, and my idea of building a compiler backend, we put that together and, and we, uh, built an SDK to launch the company.

So that is the origin, and that is the reason is, you know, AI cannot always be running in the cloud. And I'll, I'll I'll end. Uh, this, uh, part, right when I went from my first fundraising, right?

We are a VC funded company. My first fundraise, our, our seed round and, uh, the seed round was invested by Steve JSON from Future Ventures. Steve has backed Elon, uh, companies like Hotmail and, uh, uh, you know, Commonwealth Fusion, all leading, uh, edge companies in, in, in this, uh, respective industries, right?

When I pitched to the group of investors that day, I pulled out my iPhone, I called Siri, asked for time. Siri promptly answered me the time. Then I put the phone on airplane mode, called Siri again and ask for time.

The response was, sorry, Jags, we are not connected to the internet. So when Siri said, you're not connected to the internet, and I turned to the investors and I asked them the information, the time is available locally on the phone, right? And the command is very small.

Why can't we have selected commands and selected intelligence available locally on the device? Why does it have to always go to the cloud? This is the reason we need edge ai.

And that secured my first seed round. Good For you. What a great start.

I'd love. Look, I unfortunately, well, not unfortunately, I've had the pleasure of, of doing several, you know, venture backed startups and, you know, you always got that moment where, you know, you got 'em right, where the VC said, this is something, this is a must have, this is important, not just nice. Um, so here, here's my take on it.

I got kind of three different areas. I, and by the way, we're gonna come back. I want to tell people, you know, late in AI's website, how they get involved product wise and all these things.

But three things. Number one, in my mind, I, I have this like diagram in my head of I've got my core hypervisors own data center in the cloud, probably right? Then I have edge computing on the edge somewhere.

And that, that takes a lot of d everything from like a, a container, not a docker container, like a aircraft container with equipment by a cell tower or something. Or you know, somewhere along the edge, some computing resource and then on device resources, which to me is the edge isn't the edge, you know, it's kind of mushy right there. Yes.

But we're here talking really about on local device computing or not, Not. So we introduced the term called the edge continuum, right? I will, I will take you back to your chemistry class in high school.

Do you remember the distillation tower hour crude oil gets heated up and you basically go through different layers and you extract products out of it. I was reading, uh, the Economist article about data is a new oil again in mid 20. Uh, I think the article came in 2016, right?

So that is when like, you know, some synapsis pied inside and I started to work on that edge continuum then. Okay. So the way we position it is that data gets converted analog to digital at the central level.

From there, data travels through different networking nodes, different computing nodes along the pathway, either to the destination in a hypervisor or whatever that destination may be. Travels through that. Along this layer, computing is distributed heterogeneously, okay?

Our thesis is that how do you define AI for this heterogeneous environment? So you extract the right level of intel along the pathway. Okay.

So, uh, um, we have done extreme, uh, extensive business in the depart Department of Defense. And I, uh, co-authored a paper with, uh, retired vice chairman, joint Chiefs of Staff General Cartwright, uh, for the Atlantic Council, which is a, a, a leading think tank in, in defense. Familiar.

We co-authored that paper and we actually applied a similar strategy for defense, whereby we are calling it the, uh, uh, uh, tactical edge, operational edge, command edge and strategic edge, right? Four layers of the edge. And how do you process information accordingly?

I love it. Yeah. Is there a diagram or something of the edge continuum on latent ai?

Yes, there is. All of it is. I, I would recommend, you know what, to our audience go, we should get this outta the way.

ai or is it latent? com. com?

Yes, That is correct. Okay. com.

So, so Jags, let me ask you the next question then. You know, I was reading an article today, a new survey out 75 or 74% of organizations say their AI initiatives are being delayed because of access to G, right? We need these GPUs, which I don't, I get how good GPUs are, and I understand their role in ai.

I just don't know if I buy into the whole 75% of AI stuff is being delayed because of access to them. Um, but that being said, one of the issues is a lack of that kind of horsepower at the edge in on devices. Now we are seeing a I PCs, the so-called a I PCs, we are seeing, you know, apple and Google Android phones, uh, you know, with AI built in, and they've got, you know, the M four chip and all this kind of stuff, not the M fours from a Mac, whatever the latest chip is on the iPhone at eight 20, whatever.

Um, what about do we have the horsepower along this continuum to truly perform the kind of AI that, you know, the AI we want? Let's say the a you know, the, the kind of stuff that's game changing. I'll give you one example, right?

This is something that we've already done and, and we actually demonstrated with the telco. We talk about the edge continuum, right? Uh, we did a doorbell as an example, okay?

A doorbell today, if you think about it, it's motion activated. Anything that moves in front of it, you are gonna get a notification. It could be, I had a buddy of mine, uh, he had a b that was trying to enter the camera, right?

So he got 800 some notifications and on, oh God, Imagine the, the frustration there, right? So let's say we build a low, low power battery operated camera, doorbell camera, right? It could be a forbit processor.

We actually built it on a forbit processor, and the only algorithm that's running on that FORBIT processor is a human non-human detection. Is the object in the frame a human or not human? Okay.

So now once a human is identified in that frame, then the next layer of compute, it could be a set up box in your home, or it could be the MEC environment, let's say a multi access edge compute unit from a telco perspective, it could be a 5G connected, uh, um, device. So that server could be running facial recognition at that point. Is that Alan at the door?

Is that jags at the door? Do they have permission to come in? What, you know, what is the next set of rules that you need to follow?

You could do that. Let's say you're not able to identify that person. There is a human, but I don't know who that is.

Then we can pass that information. That same frame can be sent up to the third layer. Could be a CDN, could be a little bit more, uh, uh, compute heavy, uh, Part of the continuum.

Part of the continuum. And at that level, you're looking at it and saying like, okay, is that person wearing a uniform? U-P-S-U-S-P-S, FedEx?

Is that person wearing a hat? And, and, and, and covering their face? Is their face occluded?

Is that a security issue? Right? Third one, is that person carrying something in their hand?

Is their hand hidden? You know, is it a delivery of something, right? So you're able to identify in your career, able to run these d different models at that level.

So the Math, I, I get it. So it's almost a just in time system where, you know, if, if you have the horsepower, if you view this continuum Yes. Bottom to top, let's say right from the extreme edge all the way back to core, you, you, if you have the, you, you make the decision where the, where you have the horsepower, and if you don't have the horsepower, it gets kicked back one until, you know, finally at the floor.

But you only have to do it at the core if that's the only place where it could be done. Exactly. So it's alm it's like efficiency built in.

What a, that that concept alone is the money, right? I mean at, at some level. Exactly.

Right. And we are building, putting the logic in place for that. Exactly.

We are building the tools to so that you can rightsize that logic, rightsize the model, and you can deploy it on these different hardware. We are on hardware, hardware agnostic and model agnostic tools. So we are, I get it.

It's just a question of building that just in time system. Exactly. I love it.

Um, you mentioned you do a lot of work with DOD. Uh, how, how long has late name AI been in business? We've been around for six years now.

Oh, so you're, you're an old tire. Another, another overnight sensation. Exactly.

Off. Good for you guys. And I know you, you've, you've got some tremendous talent over there besides yourself.

You know, you're an accomplished Yeah. Uh, that peron yourself, but what about the rest of the team, Jacqueline? Oh, um, uh, I am blessed to have an, uh, extremely talented set of individuals packing me here.

Um, I would say my, my CTO, right? He's a celebrity when we go into AI conferences, he's, uh, he's a well-known chair for, for different events. So Sec Cha is my CTO.

He, he's amazing. My head of compilers is a guy who did his postdoc work with Dr. Bill Dally.

And Bill is the father of GPUs, right? And, and a chief s at, at Nvidia Lardo worked personally and closely with, uh, uh, with, with Bill at, at Stanford. So Abla is helping run the compiler show for us.

And then Jan, um, our ML guy, he comes from Siemens and, and, and, and a historic career from Germany to the us and he is well known in the ML and CV space, computer vision space. And he's pushing us into different, different sectors that, that we've never even thought of and, and drawing results, uh, crazily, right? So, and then we have, uh, almost close to, uh, 15, 20 PhDs in the organization.

Uh, we are growing continuously and, and we've doubled our size, uh, last year, grown over about 200% in revenues, uh, uh, year over year. So it's, uh, What a great story. Seemly lucky, uh, uh, uh, to be in the space.

What a great story. Congratulations to you for putting this all together. It, it's a, it's a feel good story.

I like it. Thank you. Hey, we're about outta time, but first of all, again, congratulations on, on the success.

I, I love what you're doing. Do come back, keep us posted. That's, and, and best of luck to you and latent ai.

That's fantastic stuff. Jag's kind, co-founder and CEO of latent AI here, ONT Techstrong tv. We're gonna take a break.

We'll be back in just a moment.