AI at the Edge: Justin Schneck on the Future of Embedded Devices

Speaker A: Foreign. And we are back. Welcome back to over the Air AI Connected Devices and the Journey. My name is Ryan Prosser, CEO of very. Today I'm joined by my very good friend Justin Schneck, co founder and Chief Product Officer over at Predeo. Justin, thanks for being on the show.

Speaker B: Hey Ryan, great to be here. Great to be back.

Speaker A: All right, so uh, for folks that don't know, give us a little plug on Peridio.

Speaker B: Paridio uh, is a embedded product platform really to help you across all the aspects uh, of the journey of creating an embedded product. Uh, we provide over the air updates, we do device management services and our fleet product. Uh, but we also have uh, made a big push into uh, showing what we believe are some great foundations towards how to structure and create uh, embedded products with our uh, OS that we're releasing, uh, coming soon called Avocado. Uh, and uh, that's going to really edge uh, end to end, help uh, across all aspects of the uh, product development journey.

Speaker A: What can we glean about Peridio's culture that you landed on Avocado?

Speaker B: Uh, that one has a little bit of a playfulness to it. Uh, it's really that we found that uh, there's a lot of companies that are building embedded products who don't want to have to care about the hardware. And uh, it's a really great way to be able to tee uh up framing on uh, Edge devices themselves. There's a wide variety of them but the ones who really just don't want to have to care about the hardware. They just want to be able to deploy, deploy maybe some machine learning models or some compute to a physical appliance. Uh then uh, for them uh, I almost uh, would say it's similar to like how we have software as a service SaaS. Well now there's hardware as a service or HAAS, hence HAAS avocados. Hence Avocado. Yeah. So it's uh, to really give you that hardware as a service feel so you don't have to really care about all of the complex parts of uh, embedded development and focus on uh, your business logic and making your product do exactly what you wanted to.

Speaker A: I did not think that you were going to be able to pull that thread together. I thought for sure you'd lost the plot. This was going nowhere. Okay, I got it. Haas Avocado. Um, now longtime listeners of the show will of course know a bit about Peridio. So uh, I mentioned at the top of the program that uh, after I think a one year hiatus, I'm back as host of the show but we had ah Peridio CEO and your co founder Bill Brock hosting the show for the last several months which we're really concerned for. Um and we've been hard at work over here at Very pulling off a pretty non trivial pivot that uh, that I wanted to make sure we got right. And um, and I asked you to be on the show as guest number one for V2 of the show as we kind of talk through that. So you know very for a long time has been focused. So for folks that don't know I'm CEO Very Very is a product development company specializing in uh, in connected products and um, you know just looking at the landscape we we've been doing AI for years, you know definitely predating all of the modern buzz around it. Um but you know in a post GPT world um, and also kind of a, there was a bit of a Moore's Law uh component to this too where like the things possible at the edge were far beyond what you could even do three years prior. And so we really tool retooled around this idea of edge AI. Um and then of course you Justin my old and dear friend uh were talking about this and um, you know sort of talking about some of the next door neighbors to edge AI and that there's, there is kind of a, I think you use the word gradient to the left and a gradient to the right as uh, you know edge AI let's say to the left kind of starts to turn into cloud AI. Um and you know as the gradient goes to the other direction embedded AI um and you know so that's kind of our topic here today. So I, by the way I'm so rusty I forgot to mention the topic in my intro. This is a core part of what I do here. So folks we're going to be talking about the gradient uh the spectrum if you will from you know cloud AI, edge AI kind of sitting in the middle embedded AI. So give us kind of like World according to Justin on that spectrum.

Speaker B: Absolutely yeah. AI uh is uh one of those things that was really born out of academia, it spent a lot of time in academic circles. Um uh there was a lot of uh, uh theory and speculation uh and practical application for long periods of time. Uh but uh more recently we've seen massive growth in the availability of really uh powerful AI tools. Assistants. Uh now we go into meetings and we see usually one or many uh note takers showing up alongside of you uh, who are there to be able to do Some great services. It's, there's some fantastic routines and workflows that come from being able to take uh, you know, really accurately textualized meeting notes, even forward feeding Those into other AIs to be able to process, you know, what is it, what did we talk about? Give me some uh, uh outlines that I can circulate, uh, what should I follow up with. And a lot of that AI is uh, really the cutting edge pieces of it and it takes a lot of computational power. We have service providers that are out there that are well known name brands these days. Even ChatGPT and the Googles of the world are all pushing forward on these fronts and they offer uh, these services with uh, just massive amounts of compute that's uh, behind them. I mean so much that uh, we're considering firing up more uh, shutdown decommissioned nuclear reactors to be able to power them. Right. Uh, that's all on the cloud AI front. And all of that raw power is really like, it's the first expressions of like where's the edge? Where to overuse that word even. Where is the like uh, the innovation uh occurring. And it's really pushing forward there.

Speaker A: Talk to me about hype cycle. Like we hear a lot about like the trough of disillusionment is one of these fascinating things that 15 years ago I think was such a niche concept and now everyone looks at it as a core framework for everything. You know, so it's like oh, where are they at in the hype cycle? Uh, what's your take on like these three, I'll call them technologies. But like I, I know that's more amorphous than that but you know, like are they in the same place or you know it seems to me that they're not at all in the same place. What, what do folks at home need to know there?

Speaker B: Yeah, it's real uh, there's a lot of uh, hype that I think is born from just the amazing uh, outcomes that we're getting that we can see from, from the practical use cases and the most high powered high performance domains. The just the areas of this and that have moved from academia moving so quickly and producing these, these fantastic tools like are showing great promise. Um, but almost through like a fun trickle down AI economics way, uh, as that stuff starts to get like uh, proven is where it starts to become more sound, to get pushed further and further down into embedded. And so uh, you know all of the, all of the like fun and, and and hypey uh work that can go on in AI is Really it's still there, it's still happening, there's still progress being made and there's a lot of people that are attracted to that. Uh, but the real value I think we're starting to see now is, is having it uh, be robust enough that we're willing to embed it into electronics. And um, you know all these products that we have today are made from electronic components and you can, you can just gloss over what they are. Really it's just software that's been cast uh, into stone, literal stone. Right. And uh, um and it has trade offs because it's uh, it needs to be miniaturized so that people uh, who are going to build a product can select your components to solve some problem that they have that their product might, might face. Right. Uh, one of these as an example is actually like a bathroom sensor, right? There's uh, occupancy sensors. It's really dumb device, they've been around forever but uh, there's a quirk with them and that uh, some of these sensors uh, need to be calibrated for the environment that they're in. And what if you're putting it into an environment that's got a lot of reflections or a noisy one. Uh and uh, in order to get past this calibration moment a lot of sensor manufacturers these days that make these components uh, started reaching for miniaturized models, AI models that they can embed into the products uh, that can satisfy uh, this sort of self initialization sequence. Um because you know AI models are great to be able to do like uh, predictions on how they should uh behave based off of the inputs that it's given. And so it's really a great domain to be able to use for uh, sipping something that on the, on the, on site uh, doesn't need like hand calibration or uh, the user to interact or interfere in a way for uh, to give it like a domain specific behavior. Uh and so I really see like yes the hype is real. Uh, there's a lot that's really being pushed forward and like the, the, the latest and greatest features of these massively powered cloud enabled AI projects. Uh but at the same time um, uh the boots on the ground, the uh, sort of unsexy portion of the work is really where I think a lot of the traction is being made in the market and um, solidifying real world use cases and starting to ground the theory from academia, uh into actual business uh cases that are going to help.

Speaker A: I want to go back to something that you said A moment ago, um, you know, it seemed like you were kind of suggesting that, um, technologies and use cases and so forth. Like I think, I mean just to visualize it and there's. It could flow the other way. But like, I think of, you know, in my mind, cloud AI on the left, edge in the middle, embedded on the right. I only think of it that way because it seems like what you're saying is like the technologies and capabilities really start on the cloud side and flow through not exclusively, but like they're harder to implement the farther you get, quote unquote, to the right. In that spectrum. Does that seem like if I'm tracking with you, that was your view?

Speaker B: Yeah, yeah. And really it's just, uh, the hardware, you know, as we shift from left to right, we're shifting from general purpose hardware, commodity hardware, to a more specialized domain. And um, this shift has been fueled heavily by the silicon manufacturers who provide the uh, heart and soul components that go into embedded products these days. Around 2020, ah, there was a big uh, shift that we started to see where uh, processors that were coming out on the market from these manufacturers, they started to have more specialized processing cores that were put into them kind of like in the early 2000s when um, computers stopped getting faster and instead you just got more CPUs, you know, like, they're like, oh, it's got eight cores, 16 cores, you know, 24 cores. Right. Similarly, we're seeing the same kind of tectonic shift in embedded processor. And um, because that complexity increases, so does the complexity on how do I get my stuff to be able to run on this, on this, uh, this new landscape, this new domain. Um, it's uh, no longer as simple as just firing up the tools that uh, run perfectly on the commodity hardware. It might require some transformation in order to get it into this, to this state.

Speaker A: So yeah, it's interesting. Like it reminds me a bit of, uh, as dumb as this will sound like, road bikers and mountain bikers will recognize the same model occurring with transmissions on bikes. You know. So like there's different group sets on these bikes. Um, so road bikers, you know, every, I don't know, Dura ace is like what all the Tour de France guys are using. Then there's Ultra ultegra, then there's 105 and you'll see for the hardware. Although now they, you know, they do run basic firmware on these transmissions and road bikes, which is fascinating. Um, uh, I think the Di2 electronic shifter has now worked its way all the way down to the 105 group set, which is kind of the, the you know, Average Joe group set. But you know, they'll, they'll start with something on Dura Ace for racers only and it'll slowly make its way down to, you know, the more mainstream. Um, this isn't necessarily the same model. I wouldn't call it embedded a high mainstream by stretch. But it does seem that you see things debut on the cloud side um, and maybe work their way through. What's your take on. You know, we just saw in the news we're recording this episode May 2025. Um, we just saw in February, uh, my very good friend Brandon McBee's company, uh, Coreweave IPO. This would be a good example of uh, of cloud AI, uh, maybe a name that not everybody knows. Am I thinking of that? Right?

Speaker B: Yeah, yeah. They're doing some great work to enable uh, uh, processing facilities that kind of break the mold of the clutches of Nvidia and their grip. You know,

Speaker A: I thought that, my understanding was that they're more like hand in glove with Nvidia. But are they, is there a frenemy situation there?

Speaker B: There's a, there's a lot going on in the industry right now. I think that there was even some recent news that uh, Nvidia was going to work together with Qualcomm, uh to do data center chips, um, that were on site with their uh, HS100 GPUs to be able to have higher performance, um, and localized compute attached to those nodes. So there's just so much going on in the industry, uh, um, that enabling the data center race for AI is huge. Um, um, uh, but I just don't. Part of that is just hard to be able to see. We were already before the AI data center race encountering issues with large uh, scale data centers themselves. Just not necessarily running out of space to put servers but running out of cold air. That was like the problem. Right. And, and there, there were a number of solutions that, that were presented. Uh, I believe Facebook built some facilities just where it was cold, um, to. And just open up, open up the windows. Right. Uh, and uh, and, and then uh, I think Microsoft, uh, uh was uh, toying with an idea of building shipping uh, containers, building racks into shipping containers and submerging them um, into the sea. Right. And I think you have to stop and ask yourself like, like how do you get here when, when you decide that you're going to boil the oceans for um, a purpose. But uh, yeah. And so it's, it's Interesting that there's this, that now this reinvigoration of like, oh, we forgot about the fact that data centers have the scaling problem, but we really enter edge AI. That's where the, the big shift is to try to, let's try to get it out of these centralized facilities. Let's make it performant. Uh, you know, we've kind of learned these lessons from blockchain. Like uh, we, we need a, we need a balance, we need to strike a balance between performance and uh, the actual product. Uh, you know we having the product is great, but if it can't, if, if it's, if it's going to require so all, all of this energy and waste, uh, then it's, it's uh, it's not necessarily going to be a very sustainable fit. So I, I want to. Before we move on. I know you're a Shimano guy and I gotta admit something. I'm a shram guy. Right? And I got, I never understood you people.

Speaker A: I never understood you people.

Speaker B: And I, I uh, uh, I had this, I had this experience that I mentioned. I mentioned. I do. I, I, I deeply focus with, with uh, my company on, on quality of software updates. And, and one day I just had, I had to get out and ride my bike to get out into nature and get away from the, the, the day to day problems with this stuff. And I, I got on my, I got on my bike, I fired up my Garmin and uh, it wouldn't let me drive until I firmware updated my trade. And so here I am getting pulled back in.

Speaker A: Imagine, you know, it's race day, you've shipped your bike out, a firmware took place during the shipping process. You're on the, you know, whatever transition between whatever, uh, swim and bike. Get on your bike and you're bricked or whatever. Not bricked but you know, you, you have to, you lose eight minutes to, to whatever. Here's my problem. Um, and I had the same problem with Nvidia. Words that refuse to put vowels where they should go. And just like SRAM is not. This is not a word that I'm comfortable saying over and over spelled in a strange way. And that was enough for me. That was to like uh, for shamano to make a lifelong follower out of me. It's spelled in a conventional way. I look at it, I'm 100% sure how to pronounce it. I have too many difficult decisions in life already. Nvidia. I have no choice. They've eaten the world. So I'm Gonna have to live with that one. But I like words that I can look at and immediately understand. This is how that word is spelled, uh is pronounced.

Speaker B: Now, uh, I know this seems like a tangent, but all of us to be able to say there's actually some great edge/edded AI products out there. And one of them that I really love is um, I have the, uh, I have the Garmin, uh, Varia, uh, uh, tail light, uh, for my setup. So I have a Garmin head unit gps. It's on the front of my bike. I've got the, they have a Varia tail light. And um, uh, embedded into this is uh, it uses what's called millimeter wave radar. And it's constantly scanning, uh, my surroundings behind me, what's close to me. And on my Garmin head unit it'll show me when cars are coming up. And so, you know, here we have something that's processing lots of data quickly and it needs to be able to make assertions on what, what, what according to its radar brain. You know, like a. Imagine, uh, just like a point cloud, like dots in space, like it thinks looks like a car, right? So it's running this little model and uh, it's making these assertions and it's, it's, it's a, it's, it's such a, it's just a delight because you know, I, I don't have to be, uh, I don't, for two things. I don't have to be like weary about when a car is coming up. I can tell when it's happening. It makes a big audible sound and I don't have to look like a dork with the rear view mirror. You know, always uh, trying to be able to keep an eye out for the surroundings. Um, I haven't had any false positives with it. It's been a delight. So there's all this work that's going on in AI, uh, in the cloud. Like we can train stuff up there. Feel free to be able to use all the horsepower you want to. I prove that a product can work. Um, but then like to, to, to make it efficient that it can get into this package that's uh, uh, you know, not that expensive. It wasn't, it wasn't crazy expensive. It wasn't outlandish for the, for the value that I'm getting out of it, especially the safety that I'm feeling out of it, uh, to get that into the consumer market and produce products we've never seen before. That's amazing. And uh, you know While there's a lot of effort, work that has to go in from getting, from the cloud experiments that you've created so that you can just get that proof of concept, get it running all the way down to actually making that final product, there's um, that, that time and effort and energy wouldn't be possible without all of this academic research with all these people, without all the hype, without everybody so invested in it. So I, I really think that there's some great stuff we're going to see come out of this.

Speaker A: You know, you, you mentioned the academic research and we're, we're at time, so I'm going to wrap this up here pretty quick. But the when, when the AI hype really took off in like 22 or so, um, and you heard a lot of rumblings, okay, this is the, you know, blockchain, you know, kind of here we go again with something that's going to like go way up and go way down. And it was really the academic research, like the mountain of academic research sitting under AI that really, you know, blockchain and other technologies didn't have. You know, I mean they went up, uh, they went down. But like there's just this giant ocean of work that has taken place in AI over the course of several decades that, you know, this was a big part of the retooling that we did at very, you know, as I'm looking at this and I'm saying this is not like other hype cycles. The, the amount of work that has gone into this over many decades is different. Um, it's more of a coming of age, like when the Internet was finally, you know, ready to uh, to go mainstream. This felt much more like that than it felt like, you know, some of the, like I said, blockchain type hypes where the use cases weren't there, there wasn't provable, roi, you didn't have this ocean of previous research and so forth. Um, so I got one last question for you. Take us, take us forward a few years. You know, it's 2030. We're looking backwards at the embedded AI world. What are some of the, uh, things that you would report back from that place in the space time continuum to now and say, hey guys, guess what? Here's a couple of things that are definitely true.

Speaker B: I think that uh, embedded AI will start to really reach sustainable growth when we get to the stage where high quality middleware tools are available to do all of the hard parts of gluing things together similar, uh, to every sort of hype cycle. On the other end of things, the sustainable growth and product, um, adoption really comes from the place where we've all landed on quality, uh, tools that we want to start building. In the early 2000s when the web bubble hit, uh, after it burst, uh, arguably Ruby on Rails came through and gave birth to tons of companies that made it easy for them to develop products. And that was where Silicon Valley really saw its, uh, you know, slow and steady climb towards sustainability. Uh, so we're just going to start to see that as well when it comes to, you know, out on the other end of this, we start to be able to get high quality tools and middleware that's going to help us, uh, build these things, um, in an approachable and reasonable way.

Speaker A: All right, well, Mr. Justin Schneck, thank you for being on the show and being guest number one of V V2. Always nice to have a friendly face.

Speaker B: Thank you so much. Ryan, I'm gonna challenge you to some sort of race, uh, on, uh, on your bike. At some point we'll go shrimp versus salato.

Speaker A: I love it there because, of course, there are no other differences between us. For folks that don't know, Justin is approximately twice as tall as me.

Speaker B: That's twice the body weight.

Speaker A: That's right. That's not necessarily an advantage. That's right. That's right. All right, let me, uh, let me close out the folks at home, uh, and thank you for listening. Okay, that was the most awkward outro possible. Uh, but we'll pretend like it was smooth. Thank you guys for listening. Join us next time. We'll be smoother next time as I meet with another AI executive and talk about things that went wrong on a journey that went right.