What's Up with Tech?
Tech Transformation with Evan Kirstel: A podcast exploring the latest trends and innovations in the tech industry, and how businesses can leverage them for growth, diving into the world of B2B, discussing strategies, trends, and sharing insights from industry leaders!
With over three decades in telecom and IT, I've mastered the art of transforming social media into a dynamic platform for audience engagement, community building, and establishing thought leadership. My approach isn't about personal brand promotion but about delivering educational and informative content to cultivate a sustainable, long-term business presence. I am the leading content creator in areas like Enterprise AI, UCaaS, CPaaS, CCaaS, Cloud, Telecom, 5G and more!
What's Up with Tech?
Transforming Developer Velocity: How CodeMetal is Revolutionizing Software Development
Interested in being a guest? Email us at admin@evankirstel.com
The gap between AI research and deployable software has long been a bottleneck for innovation. Peter Morales, CEO of CodeMetal, is bridging this divide with an approach that's fundamentally different from the AI coding assistants flooding the market.
Drawing from his physics background and experience developing AI systems for the F-35 fighter jet and drone defense systems protecting Washington DC, Peter recognized a pattern: whether in defense, academia, or Big Tech, translating high-level algorithms into hardware-optimized code was universally painful and slow. While companies like Microsoft were building AI "copilots" to help developers write code faster, Peter saw an opportunity for full "autopilots" that could handle entire development workflows with built-in verification.
What sets CodeMetal apart is their focus on trust and traceability. Rather than simply generating code that might work, their system creates comprehensive test suites and validation artifacts that prove the code meets specifications. This approach is particularly valuable for industries with stringent compliance requirements, where organizations are understandably hesitant about deploying AI-generated code in production. By starting with trusted reference implementations and providing guarantees about the generated code, CodeMetal has helped partners reduce development cycles by approximately 50%.
The challenge now isn't just technical but cultural. As Peter explains, "The shift in moving to AI tooling is actually more of a cultural shift than technology." Organizations built around traditional development processes must reinvent their workflows to fully benefit from these capabilities. For entrepreneurs entering the AI space, Peter's advice is clear: double down on your unique strengths rather than trying to compete with generalized AI systems. The companies that will thrive are those that solve specific pain points in ways that general-purpose AI cannot.
More at https://linktr.ee/EvanKirstel
Hey everybody. Really interesting chat today, as we talk about the role of AI when it comes to developer velocity and software development in general. So much going on here. We have a true innovator and expert with us, Peter from Codemetal. How are you?
Speaker 2:Great Thanks for having me, evan, excited to be here, are you?
Speaker 1:Great Thanks for having me, evan, excited to be here. Well, really excited to chat. This is an area that's generated so much enthusiasm and interest. Before that, maybe introduce yourself your journey and what was the big idea behind.
Speaker 2:Codemetalai. Yeah, my name is Peter Morales. I'm the CEO of Codemetal. I started out my career with a degree in physics and started really wanting to go into industry more than academia, and mostly because I loved coding. I got to watch a lot of the fundamental things that we were researching come to life. So you get to do first principles but then actually build some code out of it and went over to BAE Systems in Burlington I'm in our office in Boston, so not too far from here and started working on different RF and EW systems so those are radar systems and electronic warfare and part of the work that we did there was basically taking high-level ideas so things that we would write in MATLAB or or, if you're familiar with Python, popular for AI and bring that down to code that would actually run on hardware. Those could be FPGAs, those could be GPUs, those could be small embedded devices. The thing that we did really well was we actually automated a lot of the work that could be automated. So we actually generated unit tests. We generated data for those unit tests, all from this high-level code, which really made our lives easy. We would basically take then the concepts at the high level and map those down into different hardware ways of thinking, and so that was the task of basically bringing AI to the edge or different sort of signal processing algorithms to the edge, and we got to work on systems that ran on the F-35, systems that ran in the national capital region a whole bunch of really cool work that actually the F-35 work in particular, got me excited about AI. It was one of the first reasoning systems that actually ran on that airframe and it was sort of hey, this is code that can think.
Speaker 2:And that's when DeepMind had also published their Q-learning paper and that made me want to kind of go all in on AI research and colleagues recommended I go to MIT Lincoln Labs so I could take courses at MIT, there could work on different AI projects, and I ended up moving over there and so over at MIT Lincoln Labs focused on AI research with some great colleagues at Stanford, focused on computer vision projects locally in the counter UAS domain, and you know, again, it was the same sort of idea where you're working in a high level language to do the research, maybe a small team, and then it takes a big sort of group of software developers and, you know, sometimes low level hardware engineers to bring that into an actual product and we worked on deploying different prototype systems that protect the DC airspace from drone attacks and that worked. Because the lab at the time was looking to do more AI research. They farmed an AI group out of the CTO's office and I got to be one of the founding members there. So it was a really cool opportunity to work with campus on Air Force problems that the Air Force wanted us to work on and so I worked with like Davaman before he came head of media lab over there, got to work with some of the different um staff there on really interesting problems, but it was a little bit more of a paper writing exercise and I love to build and microsoft was doing the hololens and so I ended up leaving, uh, the lab after that, moving out to washington to work on device and I think the big thing and really the inspiration for CodeMetal was when I was going over to Microsoft.
Speaker 2:I had this expectation that us getting our algorithms to edge was going to be really different. I was in defense, which maybe doesn't always get the credit for software engineering, and then I was at the lab, which definitely didn't have the software engineering chops, and so I thought the actual bringing these algorithms down to hardware you know whether it be a robot or some medical device or anything it was going to be a lot easier, specifically on the HoloLens and it wasn't even though they make their own chips.
Speaker 2:It was the exact same.
Speaker 2:In fact, I still give kudos to the BAE group there Alpha Tech guys that they were probably the best that there were at bringing these algorithms to edge.
Speaker 2:And it was the same very stressful process to get these things running on hardware. And that was really the motivation for what we're doing that if this exists in every industry bringing a hardware product, tackling all the different low-level SDKs and firmware issues that there's a really good spot for automation here and sort of filling in a gap that right now real-time system developers do with some tools and specifically AI tools. And so we came up with the idea of essentially integrating large language models with formal methods and testing and other sort of traditional code techniques to right now and I think vibe coding is kind of like the hot thing that everybody's talking about. Imagine now being able to actually trust the code that comes out of that. So generating tests, generating different sort of proofs that basically show that your intent was carried out the way you wanted it to. And that's our focus at CodeMetal essentially automating a lot of the nitty-gritty processes that can be automated with full sort of autopilots rather than copilots.
Speaker 1:Brilliant, I love the mission co-pilots Brilliant, I love the, the, the mission, so big picture. Ai co-pilots if we can call them that, no offense to Microsoft, but just generally are becoming very widely used across all kinds of developer environments. So how do you see yourself, uh, on this landscape, vis-a-vis the current wave of co-pilots and other LLM-based dev tools?
Speaker 2:Yeah, so you can think about the co-pilots as really doing two things. Well, so they're positioned for clean sheet problems, meaning like I don't have any third lay down but I want to generate, let's say, like a rough prototype, and I think they're great for that. I think they're a really awesome tool for that clean sheet problem you can spin up. I mean, basically you see a bunch of people like making tiny little projects, like oh look, how quickly I got a kind of working thing, or maybe it works to their intent all the way.
Speaker 2:I think where we fit in is if you look at like sort of a production chain, like an actual. You know the giant teams that are sort of operating from companies like Toyota and automotive to you know, l3 Harris is actually customer Mars and defense. These require a lot of different expertise, meaning like they need the folks to maybe write the initial clean sheet problem, get the feature that they want, but now it gets handed off to somebody that understands real time system development, which might get handed off to somebody that knows hardware. For example, in their use case they actually do signal processing algorithms in MATLAB, move it to C++, because it's like a real-time language that they can kind of wrap their head around how it would get deployed and then actually do another copy to VHDL, which is the actual FPGA programming language that it runs.
Speaker 2:The whole process is actually something that you know, is happening all over the place, and so being able to validate from the starting point sort of the initial code that you write, using that as a reference, we can actually give sort of a provable output that's optimal for the hardware that you're targeting, and so that speeds up development times in a huge way. So where we fit in the pipeline is sort of yeah, use a co-pilot to sort of help you be 20% faster, whatever, get the clean sheet out quicker. But it's not posed from an AI perspective in the same type of way that we pose our AI problem, which is given some known good, how do I work my way to a certain target language and imbue it with expert knowledge on the hardware?
Speaker 1:Got it. So let's talk LLMs. I mean, it's amazing. The improvements are fast and furious, but there are still trust and explainability challenges, as you know better than I. So how do you think about transparency and traceability and them making code-level suggestions and those sort of issues?
Speaker 2:Yeah, yeah. So it's really interesting from again if you're pairing it with ideally, an expert, right? I think if you've ever looked at some of the studies, you know somebody who knows how to prompt these things well can actually get very good productivity gains. So you know, I think that's an important aspect of it. These are only as good as the questions that you ask them.
Speaker 2:A lot of sort of co-pilots right now are focused on auto-completion. That's sort of the learning task I referred to earlier. That's their problem, meaning you're trying to predict what you want to do. You're trying to kind of look over your shoulder as like a pair of programming eyes and save you time that way. What we're posing this as is essentially I've implemented something and you know it could be pseudocode, it could be, you know, an actual implementation of Python and MATLAB which we think is fine to write, which you could have used in LLM to write, and we want to use that as a starting point.
Speaker 2:And now, for traceability and for trust reasons, we're going to generate a whole bunch of artifacts as we port that over to the software that's actually going to run in production.
Speaker 2:And now you're going to be able to actually, for your compliance needs track the code that you're actually running in production and all the sort of derivative software projects, and have unit tests, test harnesses that actually show that it matches the initial code all the type of things you need to build that trust into the system. So I think right now there's a little bit of gap between, let's say, the LLMs and getting to that point in that, like sure, I've generated the code, but how am I going to trust what came out of it? We're starting at the point of I have something I trust and I've created, but I've created it in a really sort of efficient way for prototype development, feature development, so high level languages like Python or MATLAB, and now I want to go from there to the target. So I think that's where LLMs today are going to be. You know, I'm sure a lot of CIOs and CSOs are freaking out thinking about all this sort of things being dumped into codebases and put into production.
Speaker 1:Well, that's a great point. So let's talk data, and you have all this DoD experience so you understand how sensitive enterprise or IP can be. What's your sort of philosophy there?
Speaker 2:So I think a lot of these companies they either have to get there in the trust level where they can have somebody actually own their data, but a lot of I think companies are moving to running on-prem GPTs or code models. Especially you look at like DeepSeq popping up being fully open source and, you know, making companies question should we just host this all and give the capability internally, you know, to our employees to be able to have a GPT-like capability trained on our own data, rather than serve it up to you know, especially you know the cloud hosted solution. So for us, a lot of our customers want to be able to run everything that we do on prem and we support that deployments on prem yeah.
Speaker 1:Fantastic, and so how do you work across different enterprise sizes? You mentioned you know small nibble teams and you know you've got big defense contractors and enterprises. How do you see that working across different cultures and organizations?
Speaker 2:the shift in moving to AI tooling is actually more of a cultural shift than times, even technology. It's interesting, as we're you can imagine us that are proposing hey, right now you do the process that takes 12 months, we can give you, with guarantees, kind of the same code you know, let's say, 80% you know, in a month. I think that becomes something that requires a lot of sort of rethinking of how they organize their, their organization. So it I'm really curious and sort of intently watching how all these companies react to the you know jump in sort of AI technology. I think right now we're at the point we were talking about LLMs earlier where, like, you can do some neat things but, like, once people start specializing these tools and building them into workflows where, let's say, the scaffolding for it really ties into their existing processes, that's where you're going to see these huge productivity games and that's where, pretty much organizationally, people are going to have to rethink how they build things.
Speaker 1:Wow, that's a huge discussion. Maybe a different time. In the meantime, let's talk impact.
Speaker 2:You know any examples or anecdotes or stories, you can tell on where CodeMetal is helping with that velocity.
Speaker 2:Yeah, publicly I can talk about one of the teams I'm working with, you know, on FPGA development. So they're looking at processes from 6 to 12 months and right now this is cutting them down about 50%, and so right now it is cutting them down about 50% and so right now it's not fully automated, meaning like there are bits where the sort of person needs to come in. But we're working closely with our partners at L3Harris and other companies like that that play with FPGAs at essentially giving them the trust and sort of verification outputs that let them sort of hand off more and more to the AI tooling of verification outputs that let them sort of hand off more and more to the AI tooling Fantastic.
Speaker 1:And where do you see the product? You know CodeMental, your team going In two to three years. Everything's moving so quickly. How do you stay ahead of the curve and not get disrupted by the same tools you're using?
Speaker 2:Yeah, it's interesting, I think you know really it's being opinionated about. You know where we're going and so I think we have strong confidence in the sort of verification bits and the validation and you know we invest in the things that may be. You know, completely additive meaning like we're not necessarily trying to train a giant LLM and run in that direction. We're leveraging, you know, the current state of the art in our pipeline and everything's additive to those tools. So you know we don't feel we're necessarily going to race with open AI or any of the folks building the code models. We want to basically scale with them so as those models get better, our tool runs faster, it gives the results quicker.
Speaker 2:Right now it takes quite a bit of time to generate a complex, let's say, target, meaning going from like a high level to a very complex hardware target as the models get better. That's going to speed things up. But yeah, I think I would say the organization's you know targeting sort of a time horizon of tech that's coming online in the next year and you know sort of that sweet spot of like let's always be looking ahead and pushing towards that one year out where we're not so far that we're a pure research org and we're not so close where we're going to get gobbled up by, you know, the next iteration of an AI tool.
Speaker 1:Yeah, great, great challenge there. And any advice for early stage founders, entrepreneurs, maybe folks in academia, you know, trying to build out AI first products. You've been through it now, so what advice would you give that?
Speaker 2:I think if you're an entrepreneur you know, leverage your value prop.
Speaker 2:From what do you have that other people don't? So you know our focus is the integration of compiler technology and formal methods, right, and so we're leaning on the AI explosion. We've got great people from OpenAI and other places at the company helping with the AI side, but what we're really doubling down on are the things that make us different. So if I could generate code that I guarantee certain like pain in the butt even for, like a manual developer, compliance needs to be met, that's a huge value prop and that's a huge differentiator that we're saying, hey, our code actually meets these compliance standards and has a proof. You know that you can't get from an LLM. That sort of you know relies on randomness.
Speaker 2:So, yeah, if you've been an entrepreneur, I think lean on the things that make you different. Those are the things that make you stand out from the rest of the crowd. And if you're in academia, you know for my days working at MIT, I think that's I got to imagine a lot of students are feeling scooped all the time with how much craziness is going on. So, again, I love labs with like great North stars, meaning they've got like something far in the distance that they're trying to target and I think, looking for those labs that are trying to do something different from everybody and not just figuring out what's the plus one but what's the like 10 years out do a better job at staying relevant. Great advice.
Speaker 1:What about you and the team? What are you up to over the next few weeks, months? Any travel?
Speaker 2:events talks.
Speaker 1:What's on your radar?
Speaker 2:Yeah, so this year we did CES, which was great for us.
Speaker 2:We're heading out to an automotive event in Detroit Forgive me, it's escaping right now the name of it so we'll be out there in a couple of weeks. We've just got a new hire we're going to announce. We're super excited about helping us on the growth side. But, yeah, lots going on from the growth side, lots going on from where you can see us in public. And yeah, if you're interested in the stuff I'm talking about, feel free to hit me an email and interested in the stuff I'm talking about, feel free to hit me an email and we are always hiring.
Speaker 1:Well, that's good news. So thanks so much for sharing the vision, and we'll have to grab a beer at some point here in the Boston area, peter. Thanks for joining, sharing the vision and mission.
Speaker 2:Yeah, thanks, Evan, for having me and great to meet you.
Speaker 1:Likewise Thanks everyone for watching and sharing. Take care.