EDGE AI POD
Discover the cutting-edge world of energy-efficient machine learning, edge AI, hardware accelerators, software algorithms, and real-world use cases with this podcast feed from all things in the world's largest EDGE AI community.
These are shows like EDGE AI Talks, EDGE AI Blueprints as well as EDGE AI FOUNDATION event talks on a range of research, product and business topics.
Join us to stay informed and inspired!
EDGE AI POD
How embedUR is bridging the gap between embedded development and AI with Rajesh Subramanian
The technological pendulum has swung dramatically over the last two decades. From desktop computing to cloud dominance and now back to the critical importance of edge devices, we're witnessing a renaissance in embedded systems. But this time, they're getting smart.
Rajesh, founder of embedUR, takes us on a journey through this evolution, explaining how his company transformed from connectivity specialists to edge AI innovators. Founded in 2004 to accelerate embedded product development, EmbedUR has positioned itself at the fascinating intersection where traditional embedded engineering meets artificial intelligence. This convergence creates unique challenges – embedded developers understand hardware constraints while AI engineers work in high-level abstractions. Bridging this gap requires careful training, collaboration, and a deep understanding of both worlds.
The real magic happens when we see edge AI in action. Imagine headphones that filter background noise without cloud connectivity, privacy sensors that recognize you without capturing detailed facial images, or coffee machines that remember your preferences just by detecting your presence. These aren't futuristic concepts but working demonstrations EmbedUR has created with partners like STMicro, Synaptics, and NXP. What makes these implementations particularly valuable is their independence from cloud connectivity, enhancing both privacy and security.
Yet commercializing edge AI presents significant hurdles. The journey from a 96% accurate demo to a 99% reliable product involves months of testing across diverse environments and user populations. As Rajesh points out, "Your model is only as good as your dataset," highlighting the critical importance of data curation. Through their Model Nova platform and global partnerships, EmbedUR is helping companies navigate this complex transition from prototype to market-ready product. Ready to explore how intelligence at the edge might transform your industry? The revolution is already underway.
Learn more about the EDGE AI FOUNDATION - edgeaifoundation.org
cool, rajesh, good to see you again. When was the last time I saw you? Was it in Embedded World not too long ago, feels like two weeks ago. Yeah, the whole first few months of this year is like between CES and.
Speaker 2:Embedded World and Austin. We had that thing in Austin it's kind of like nonstop Now.
Speaker 1:I feel like it's pumped the brakes a little bit, which is good.
Speaker 2:Before you even think about it, it's going to be July.
Speaker 1:I know, yeah, marty, I've already. Well, we have our Milan event in July, which I've already booked my tickets for. So it's like, yeah, the year's half over, but anyway. Oh, it's like, yeah, the year's half over, but anyway. Oh, yeah, good to see you. And are you dialing in from California, or where are you?
Speaker 2:Yes, I'm home at California. I'm actually dialing in from San Ramon, fantastic, yes.
Speaker 1:I'm in Bellevue, washington, here, nice and rainy Perfect.
Speaker 2:It's been raining the last couple of days it was sunny a few days ago, but now it's a good thing.
Speaker 1:So EmbedUR. So you know, one of the reasons we do this is we have lots of different partners in the foundation. Some of them are super well-known, Some of them not as well-known. A lot of folks know who EmbedUR is, but maybe some people don't. So I wanted to kind of just talk to you and kind of let people know what EmbedUR is up to, what you're up to, and let the audience know what's happening.
Speaker 2:Yeah, thank you again, pete. Thanks for taking the time to do this. Sure, well, primarily, I started the company in 2004, and the goal was to build embedded software for some of our customers and provide embedded engineering as a service to some of our customers so they can bring embedded products to market faster. Now, over the years, we started off as a highly experienced connectivity team and meaning working on Bluetooth and Wi-Fi, and then we pivoted towards IoT, so building IoT software, and then in the last four or five years, it's been talk about artificial intelligence. And then how do we pivot that intelligence and take it to the edge and move it to smaller devices where we have the embedded skills? But bringing intelligence to embedded devices is a whole, completely different ballgame, and that's what we've been engaged in in the last four or five years and last three years and again thanks to the Edge AI Foundation, we've been able to work more closely with your teams and progress and work with some of the partners in the group to advance their product deployment opportunities.
Speaker 1:Yeah, I think there's an interesting skill gap between embedded developers. I used to be an embedded developer with BIOS and stuff and then there's an interesting skill gap, you know, between embedded developers. I used to be an embedded developer for BIOS and stuff and then there's AI engineers, right, maybe? They're typically kind of working in Python and sort of they're up here, you know, and embedded folks are down here and so like it must have been an interesting challenge for you guys to sort of how do you combine those skills together, hire people, or did you hire people that had the skills or you trained them up or what was the?
Speaker 2:So we took a three pillar approach right. First, retool our engineers to think bigger in terms of understanding what is intelligence number one. Then we went and brought in data scientists who can analyze data, and then we brought these two together with coaching. So, basically, technology took the team. It was a lot of coaching and bringing them together.
Speaker 2:And if you think about what's happened in the last four years to five or six years, in my opinion, we went all the way from desktops to cloud right, and then AWS, azure, google Cloud, oracle all these became really, really popular. And then there was a huge. If you look at education and students graduating from computer science, they were all about, oh, I want to go work on the cloud because that's the coolest thing, and I want to do cloud compute, I want to build DevOps for the cloud. So that became a thing. But in the process, the embedded software piece was getting very less attention and, as a result, not many embedded engineers were out there experienced ones, but even kids getting out of school didn't want to go do embedded because, wow, what is embedded? I would rather do cloud.
Speaker 2:And then now we've seen the whole thing. We went from servers, then we went to desktops. Then we said no, we went to cloud. If you think about the whole edge piece, now we are coming back to the edge, saying, okay, edge is still important. And now, if you're at the edge and for smaller devices and day-to-day devices, variables and, you know, devices at home, industrial automation, devices for industrial automation, they all need edge software and embedded software. So we're in a great place right now where we're finding a lot of customers coming and asking us hey, really, this embedded engineering skill is becoming important and on top of that, the idea of edge intelligence and artificial intelligence for the edge is becoming important. So we're kind of in a sweet spot. Is what we're finding?
Speaker 1:Yeah, no, it's interesting because I think, as software developers you know people are attracted to the cloud, because the last thing a software developer wants is hardware to get in the way of their code. Exactly, and the cloud is nice because it's all virtualized. You don't even know what hardware you're running on. But when you get into embedded devices you devices you know you need to be at one with the hardware, right.
Speaker 2:Yeah, and and and again. Right, I'm I. When I went to school, I was coding in, coding with microprocessors. Right, I mean, how many kids code with microprocessors these days? Very less. Right, I mean, you need to know assembly language, right, and you need to know what your RAM is, what your flash is, how do you optimize memory, how do you optimize power? All that is becoming more and more important, and if you again think about this, this is mind blowing, right, I mean, I was.
Speaker 2:So it was Taiwan, taipei, at the AGI summit last year. So these guys approached us and you know what they're building? They're building power stabilizers. In a power stabilizer there's no active component, right, it's all passive. Now these guys were thinking of putting, I think, an STMicro one of the silicon vendors chips with an NPU on it. Now, the moment you put an NPU on it and upgrade your microcontroller with an NPU, it's a whole different ballgame. All of a sudden, they need software. It's a whole different ballgame, you know. All of a sudden they need software. So it's getting to a point where if you're building anything, you will need software going forward. So every company has to transition into a software company as well.
Speaker 1:True, true, yeah. Every company is becoming a tech company and every company is becoming a software company. Even we have Procter Gamble on. I mean, you know, I use their toothbrush and there's lots of software in that toothbrush.
Speaker 2:It's crazy, a lot of software right yeah.
Speaker 1:Yeah, so no, and I know you guys had speaking of Embedded World. You had a pretty cool booth there and it was packed with demos. It was like the Disneyland of Edge AI demos in there. What were some of your favorite examples?
Speaker 2:Some of the most favorite ones. That was we did one on audio denoising with the model right on a on an ST micro. That was just mind-blowing. We had so many customers, uh potential customers, stopped by and ST themselves stopped by multiple times. And this is, this is a game changer, because you can now build these headsets at low cost without having to have a DSP in it, with just an audio codec and a model. You think about it, it's like just completely different, and the use cases are just so many, right, I mean hard apps, people working in aviation, uh, people on a motorbike, and, and so that was. That was pretty cool, because you can still hear ambient noise.
Speaker 1:Like it will suppress the background noise, but I can still hear everybody else talking yeah, yeah and that's amazing, the um speaking of audio, we have a new audio AI working group and one of the cool things about audio AI is that you can use audio to detect things, like in a room, without using cameras right, so you can hear things you can hear, or in your wildlife or environmental monitoring. For example, you can hear birds. You can even distinguish the number of birds bugs you can hear.
Speaker 1:So this even distinguish the number of birds bugs you can hear. So this audio thing, which you know kind of was like not as cool, is now like the new thing because you can do all kinds of environmental detection or spatial detection even with audio, especially like gunshot detection, glass breaking detection, all kinds of stuff. So it's now become like the hot thing, audio ai, which is kind of cool.
Speaker 2:And the next cool thing that we demonstrated through one of our partners, synaptics, is the sensor. It's a privacy sensor but it's not really a camera.
Speaker 2:But you can look at it. It just sketches your outline of your face and then you can register with it and the moment you register it knows it's you. So then you can control. We show a person controlling lighting. So as soon as I register and I choose yellow to be my color and then somebody else comes and they choose blue, it recognizes and changes the color of the light. I mean wide-ranging applications where you can go to your coffee machine and there's a video we did about that where you incorporate this in a day-to-day coffee machine.
Speaker 1:It knows how to set the grind level, temperature level and how many shots it does the uh, it sort of does a detection based on like radio waves, or it's not a camera, or is it um, it is sort of a camera.
Speaker 2:We call it a privacy sensor. It still gets your face, but it actually doesn't register your face as such picture yeah, that's pretty cool yeah, and and so that, and so. That way you don't have to worry that it's like a full-blown camera that's looking at you all the time.
Speaker 1:It's really not. That's actually a big, big issue with privacy. These days, people don't want to use cameras, but they do want to know how many people are in the room or how long are they there and what's happening, and that's.
Speaker 2:That's another application for it, and it particularly becomes very important in Europe, where GDPR is very prevalent, right? I mean, they're very, very sensitive about no, no, no, are you registering? Is this in the database? No, none of that stuff, right? It's just facial contours that are being mapped back to the model, and then it knows, and that's all it knows. There's nothing else. And the cool thing about it is you can register up to 250 people in a small amount of memory without connectivity back to the cloud, which is completely so. Nobody can hack into your coffee machine now, right, you don't need bluetooth connectivity, right? I mean, or your thermostat for that meter, or your doorbell camera, for that matter. So, yeah, unbelievable applications.
Speaker 2:Yeah, and then we did one which was a huge hit with our partner, stmicro. We did facial recognition and image segmentation models running. We had three models running on it where you could again register yourself and it will still recognize you. Plus, it will also do image segmentation to figure out how many people are in the room. So it's like something that you can use in a conference room and that was very well accepted. And now we're getting a lot of requests. In fact, I was in a meeting with the head of AI strategy, miguel from STMicro, and he was pretty impressed, and so they've named us as one of a premium partner because we did all of this without any help from STMicro and they were pretty impressed with that Well, that's good.
Speaker 1:So when you guys do these projects, I mean you're doing it sort of soup to nuts, because you've got experience with, you know the connectivity you mentioned and the MCUs and the MPUs, and so you do like you do design, help do hardware design as well as the software implementation, or what's your.
Speaker 2:So you know, we announced Model Nova in Milan last year, right? So Model Nova is, and that's something, again thanks that we've also kind of extended to AGI Labs. And if you think of it, what we're doing today is we are going out there and picking up models, and some models we're building ourselves and actually pushing it out to Model Nova, and then we're optimizing the hell out of the model, so that way you get accuracies of 96%. It's still not good enough for productization, but it's great for you to try a small mini POC. And then our goal is that, okay, potential customers who are either chip vendors or their customers can come in and use Model Nova and, let's say they have a specific application. They can go pick one of those models and try a POC on a Raspberry Pi or whatever, and then, once they're happy with it, then it's now taking it from 96% to 99% in terms of model accuracy.
Speaker 2:Number one and number two maybe it becomes a question of retraining the model and giving them the same accuracy. Third step is move that model to a chip of their choice. Maybe it's a Synaptics chip, maybe it's a NextBeat chip. So we do all of this for our customers and that's where we bring a lot of value because we've already kind of done the groundwork. And again and you talked about this as well this was again at Embedded World, where you had that panel discussion with.
Speaker 2:Ed and Paul and a whole bunch of people.
Speaker 1:Yeah.
Speaker 2:One of the biggest issues we are finding is data sets right. Your model is only as good as your data set, and that's where we are working really hard to curate the data set, because if you don't curate the data set, junk data means junk model.
Speaker 1:Yeah, exactly, and that's the biggest value.
Speaker 2:And now take that model and optimize it completely which is our next biggest strength to low-cost, low-power environments.
Speaker 1:Right, right, interesting. Yeah, I think it's interesting. You mentioned that commercialization aspect is kind of like the secret sauce. You know everyone, I mean lots of people could kind of stitch some stuff together and get something working, but to actually bring it all the way through and optimize it and wring out all the bugs and the sustain that, you know the this privacy sensor I was talking about.
Speaker 2:Uh, pete, you believe it or not, we've been testing it for six months. You know you would think it's a simple thing, right, Right, but the problem is it's like okay, ambient lighting, dark room blinds on, blinds off different types of people, different aspects of the communities of people across the world. Right, Because this has to work world over.
Speaker 1:Right.
Speaker 2:And so it needs to be able to recognize all sorts of people in terms of lighting, in terms of, you know, posturing whatever yeah, all the conditions right yeah. And it was so hard to even get to 98%. And so the two big challenges right in this whole, it's all great to have an NPU. The two big challenges that we are finding curating the right data set and then getting 99% model accuracy.
Speaker 1:Yeah, yeah, tuning it, tuning it yeah, like you said there's a huge leap.
Speaker 1:A lot of people don't realize. When they go see a demo, usually that's a pretty controlled demo and it's like you know, like you said, the lighting is certain way and blah, blah, blah. But to actually ship it, you know, to ship it and that's where you guys come in is to help companies actually ship stuff Exactly which is like in that whole stack. We talk about the stack of edge, AI, from metal to cloud. You know, everyone plays a role. Some people make chips, some people make models, whatever, but the, the solution provider, the person that's together and commercializes it, kind of like the critical that's in the critical path. So that's what you guys do.
Speaker 2:That's the mission critical. That's where we come in. Yeah, we, we actually make it real. We really make it real.
Speaker 1:Cool, and so do you. Are you a worldwide kind of you're doing work worldwide now, or do you have particular regions of focus or what's?
Speaker 2:No, worldwide. What are the hot?
Speaker 1:industries or the hot geos.
Speaker 2:Worldwide, peter, and that's I mean I have to thank the AGI Foundation for that right and then becoming active leaders and trying to participate and understanding the ecosystem was the first big eye-opener and that led to a whole bunch of partnerships. So now, because of the partnerships we are working with, I'll give you an example we showed demos with NXP at Embedded World. We showed demos with Synaptics, we showed demos with Infineon, we showed demos with STMicro Now all of them and we also showed some demos with Infineon. We showed demos with STMicro now all of them and we also showed some demos with Silicon Labs, though it was not available for public viewing, it was very specific to them and so, as a result, they cater to the entire world. Now this has opened us up to, we believe, an abundance in terms of opportunity that we now need to evaluate and make sure that we go after the best ones that make a bigger impact for the world and the community and for us.
Speaker 1:Sure, yeah. Well, it sounds like more, uh, more miles on the airplane than maybe yeah, yeah, which is always true, that's true. Yeah, we, we tend to meet up in anywhere but our home locations, right, exactly. Yeah, cool. Well, rajesh, this has been great. I think people get in a an interesting, good picture of embed you are and what's happening, and you know people that come to Milan in July, july 2nd through 4th. We'll see you there, I'm sure.
Speaker 2:Yeah, we're super excited about Milan.
Speaker 1:More demos.
Speaker 2:We're doing something completely different, pete, so this is going to be pretty cool.
Speaker 1:You guys always have the top end For folks that don't know. It always looks really good. You always have the demos. You always have. The stuff is working really nice, thank you, we have an awesome team.
Speaker 2:These guys Eric, john Sai and the rest of the team engineering team they're all excited about this, so it's going to be fun. It's going to be completely different.
Speaker 1:Good Sounds good. All right, Rajesh. Well, I'm sure I'll see you before then, but thanks for your time today, thank you.
Speaker 2:Thanks for your time.