Episode Summary
The episode titled "AI-Powered BOM and Factory OS — with Duro and First Resonance" delves into the integration of artificial intelligence (AI) within product lifecycle management (PLM) and manufacturing operations. Hosted by Michael Finocchiaro, the conversation features Karan Talati from First Resonance, a company that develops a factory operating system for mission-critical industries, and Michael Corr from Duro Labs, which offers cloud-native PLM solutions with an emphasis on agile methodologies and API-first data management. Both companies are at the forefront of leveraging AI to streamline operations and enhance product development processes.
Key technical insights discussed include the evolving perception of AI within these industries, where initial skepticism has given way to recognition of significant cost savings and operational efficiencies. Additionally, both guests highlight the importance of user adoption through intuitive interfaces and gradual implementation strategies, such as low-code platforms like Duro's Ion Actions, which enable non-technical users to benefit from advanced AI functionalities without extensive coding knowledge.
For PLM and engineering professionals, the episode underscores the critical role of embracing AI technologies to drive innovation and improve operational efficiency. It emphasizes that while AI adoption is still in its early stages, companies must proactively integrate these tools to stay competitive, leveraging insights from leading-edge applications to inform their own strategies.
Full Transcript
Michael FinocchiaroWelcome to the AI Across the Product lifecycle. This is my your host Michael Finocchiaro And I'm so happy to have Karan Talati and Michael Corr respectively first Resonance of Duro It's gonna be really exciting to hear about them. Let's get an interest from Karan first. he's got a coffee cup. Fantastic. Yeah. Taking small sifts. ⁓ Hi, Michael and Michael. Speaking to two Michaels this morning, so fantastic to be on. ⁓ Thanks everybody for joining. I'm Karan Talati, CEO and co-founder of First Resonance. At First Resonance, we are building a factory operating system now powering something between 50 and 60 companies in mission critical industries and quality critical industries, such as aerospace, energy, defense tech, robotics. My background comes from a mix of hardware and software. mechanical engineer by academic background and morphed into making sure the right information moves from various parts of traditionally disconnected operations back at SpaceX. I would say a fertile ground for applying innovative technology, especially given all of these changes going on in the world with supply chain manufacturing, advanced manufacturing in aerospace. excited to share some of those learnings and what we're doing here at First Resonance. Awesome. I'm really looking forward to it. ⁓ Just to the audience, I am monitoring LinkedIn. So if you want to put your questions in the chat, I will ask Michael and Karan as we're going. Michael, why don't you tell us about what you do at the awesome stuff at Duro? Yeah, thanks, Michael. Always good to see both of you guys. Michael Corr co-founder and CEO of Duro. So we're a cloud and AI native PLM, product lifecycle management. And what differentiates us is really our view and embracing of the cultural shift that's happened in the hardware industry, where there's a more progressive and fast paced mentality where more more hardware engineers are actually very versed in software development and they look to adopt software development best practices into the hardware team development. And so we practice more agile approaches, more plug and play, out of the box capabilities. very API first approach to managing data. And our vision is really how can we establish that digital thread across all the different data resources, the inputs and the outputs that are needed to find a many-factor quality. And how do we let engineers be integrated and not administrators to the data platform? So I wanted to break up this interview kind of in two parts ⁓ because what's really hard, of course, right now is to separate the hype from the reality with AI, right? ⁓ And so I know both of you guys are really smart and I know that you ⁓ used AI in the conception and in the actual product itself. And ⁓ I'd like to understand first, like, What was your perception of AI before you even started? You probably already had used it with ChatGBT or maybe you had a background with ML and AI even before the open AI kind of burst the bubble on that. what, ⁓ could you guys give me an idea? Like, what was your perception of AI before? Did you see it as like, my God, or you were like, well, know, maybe it's a, go ahead, Michael. Yeah, that's a great question. It's definitely evolved quite a bit, you know, when AI started becoming into the mainstream. took a wait and see approach. I wasn't quite sure where it was going to land. think the technology itself is really mostly focused concentrated on early adopters. And so I think there was still some speculation about its maturity or how dependable it could be to actually incorporating into a business. We weren't against it or anything, but I think we were just kind of sitting on the sidelines. Standoffish? Yeah, just kind of letting it happen. We had our own business to take care of. plenty of features to focus on and great customers to work with. And so there wasn't really a much demand, but I actually, you know, for me personally, I actually had this, you know, aha moment where I guess was maybe a little year and a half ago or so I was at a conference about AI for startups and lots of people were talking about how AI technologies were helping them build companies, not so much just in the idea of putting AI into their products. And that's where I had this huge aha moment, you know, AI was allowing startups to move at an order of magnitude faster than their predecessors like us included. Like we started a company in 2018. And so first I was inspired, like, oh my gosh, you know, like we got to get our team to start using AI into their daily, you know, best practices and marketing and sales and everything. But then I had a like a fear moment, you know, don't know if we're allowed to curse on this, on this, this podcast, but I was like, oh. F, you know, if we don't embrace this, our competitors will. Right. And that was really the fire under our butt that was like, okay, this is real now. And someone with a similar concept, idea or vision as Dero starting today, this was in 2024, could catch up to us at a much faster pace than we had, you know, in our own journey. And that's what really was. the motivating factor and I literally I raced back to the office, you know, went on slack. You know, I knew a couple people in our team had been poking around with AI tools to help them other day to day. I kind of got everyone together and started like our own slack support group. You're like, okay, everybody's using AI. We don't care if it's tech, GPT, know, entropic pick your tool of choice. Let's help each other learn from each other. How can we make our own productivity better embracing AI technology? And that was really good at the moment. And now it's, it's ubiquitous. It's a must have at our company. Is it the same thing that you experienced? You know, I definitely have had a lot of the experiences that Michael just shared. ⁓ I'll add a little bit of a hybrid ⁓ experience as well. So stepping back many years, Michael, you touched on coming from an AI ML background. That's exactly what I was doing before. AI. And now when we say AI in 2025, really people are focused on LLMs, large language models, and that's how we think about AI. But AI has been around for decades, right? Research being done in the 70s, 80s. And really, actually, one of the interesting things is there's kind of this counterintuitive perception where people think that AI is more advanced than machine learning, ML. It's not. The definition of it is just human-like kind of behavior. which you could do in many ways, but no doubt, large language models are very impressive. So going back, ⁓ my experience, I kind of always knew it was inevitable. I spent a few years at SpaceX working on the data pipelines that allowed us to compare ⁓ operational data to manufacturing data. And there's definitely AI functionality, machine learning, tools that we used in that for anomaly detection and some of these advanced algorithms. So even beyond SpaceX and some of those more stochastic and statistical kind of approaches, I actually ended up at a machine learning company doing some of the same things where we were analyzing a broad set of data, correlating that data, creating training sets, et cetera. So I kind of knew it was inevitable. I will admit, though, in 2023, when the world experienced the chat GPT moment, I was also taken aback in that. I didn't realize. ⁓ just how advanced LLM research had gotten and how quickly it would move from research into commercial ⁓ applications that OpenAI made possible just through a simple ⁓ text box. Even still today, we think about just how amazing that is. When we have these moments, the internet, mobile, and now think LLMs being available for the broader public, it's pretty profound. Ever since then, we've taken shots on goal here at First Resonance. ⁓ I would say we're in our third era almost of using AI in the product. And I will admit the first two were failed attempts. ⁓ The first one was simply trying to wrap like ChatGPT like everybody else was. We put very minimal resources behind that because we had already caught on to this early kind of downplay of AI and being a ChatGPT wrapper. It wasn't going to go far and it definitely wasn't going go far in manufacturing. In 2024, there was some hype behind fine tuning models and we actually did expend good engineering effort behind there. Ultimately, the industry arrived at that being a dead end as well. But it wasn't until 2025 this year where I think a lot of things started to change. And I think we'll talk about that in a moment here, but both inside the business. how we use AI like Michael said across sales, marketing, of course, software development, which might be, again, one of these ironic places that AI is the most disruptive. ⁓ And of course, now embedding into the product making it available for manufacturers. ⁓ Going after manufacturing and serving manufacturers, just naturally going to be, we're going to have to be patient, let's say, because they're not necessarily the earliest adopters. That's really well said. It makes me think, it was kind of shocking to see ChatGPT with just that little interface, which is a text box. It's sort of like the same revolution that Google did, Just with one line, and that line changed everything. When you simplify it, it's just a good lesson in UX, right? If you simplify it and it does what you, the one thing it's supposed to do, then it's probably gonna, it's gonna work. ⁓ And it's interesting you mentioned the development thing because of course we're seeing all this stuff about vibe coding and of course lovable things you just do a line of text and suddenly you've got a fully working app covering every possible corner case. Well, probably not because I played with that too. So what ⁓ like what was your experience in using AI to develop the product? Not let's talk after that about AI inside the product but How did you approach it? Did you try vibe coding? it just, do you have an assistant and you were just using it to paste your documentation? How are you using it on, how are the developers of Duro using it to develop this new version of Duro that's AI based? Multiple facets. I think it's really dependent on the task at hand. I'd say we did do some vibe coding, you know, which allowed us to do some rapid prototyping. wasn't intentionally, it was. It allowed us to do things we hadn't done before to kind of, kind of how to thought of it was like a skunk whoops team where they could vibe code to mock some new features, quickly stand it up. Didn't have to have, you know, high quality code behind it, but it was able to at least portray our vision for what the product was to be. And we can put it in front of prospects and customers and say, how does this fit? Once we kind of, and iterate immediately on that, right? And have a pretty powerful product. And then once we learn from that experience, those user interviews, we take the best of the best of that product and put it into our main product. And in that case, it was a combination of traditional coding styles, but also using AI really to do, you know, the grunt work I'll say, you know, what AI is very good at honestly is brainstorming. You know, you're thinking of ideas like, here's how I'm thinking about modeling this or architecting this feature. It certainly, you know, with a history of know, dozens of developers over Duro's career working on the code base. It's very hard for an individual to have a firm understanding of the entire code base. And so AI is a very good assistant at like asking a question, hey, dig into this feature for me, give me a summary of how it's implemented so that I can continue that design pattern. I can go on top of it. I don't break any established architecture. And you were using this inside of like a Microsoft code? I think every developer is free to use their own IDE. I think we've all gravitated to cursor and using Claude. We found that for code development, that is the most reliable. personally, I think they're all depropping each other. And so maybe in a year, another AI tool would be better. But we're reasonably agnostic, too. think for that part, because you can establish, certainly with Claude, you can establish your Claude at MD files. so that everyone is at least using the same types of prompts and is picking up the same design pattern. So developers are free to use their own development environment, whatever they favor. But there is another thing that honestly what it allowed us to do is certainly me, so my obviously as a founder, I was the lead product manager and visionary for this whole company to begin with. And we now have a larger team that assists with that. And so my background, I didn't mention it the top of the column, ⁓ Double ECS and so, you know, I programmed my career but as I've migrated to more managerial responsibilities of that further away from code, I don't claim to be current. And so at some times, you know, as a product manager, you know, you dive into writing user stories and acceptance criteria, what have you, but there's still room for interpretation mistakes and you some cycles because a developer or you didn't, or it wasn't clear, or even you weren't quite sure how the feature would need to be implemented, so you saw it in your hands. And so what AI did is actually empowered me as a product manager is now I could actually write code production code because, you know, for those reasons, I explained as a product manager, the more complex business logic ideas, you know, common fit functionality I leave to the rest of developers and some other advanced things that I want to, you know, ⁓ discredit that, but there was some more advanced like PLM specific business ideas that I wouldn't expect someone. didn't have the same history in the industry as I did to understand it. And so it empowered me as a product manager to actually build the feature and iterate faster because then I can have it in front of me and I can play with it and be like, yes, this is right or no, this is right. And it saved me so much time from writing exhaustive user stories and acceptance criteria and then also having those feedback loops. And so that is something how I also became more productive as a product manager. I think on a separate discussion, Michael, you mentioned that Like the original Duro took three or four years to make and then you, the AI version using what you knew, but you still got the development time to what six months or something like that. Which is astounding. ⁓ Karen, did you have a similar experience at first residence? So you said you did two first attempts and then third attempt. Was the third attempt better because of AI? Yeah. So, well, just sticking to using AI for development, I would say all this has really picked up in the last. six to 12 months across the entire industry. So I'll come back to those examples of our first, second, and third wave in building AI into the product. ⁓ But I echo everything that Michael just mentioned. I actually forgot, Michael and I know each other well, and I actually forgot that Michael, you have a CS background. ⁓ I think both of us in what we have to do on a day-to-day has now gone on to other things, including myself. So just to add a few... pieces to this. I view everything going on and I think, ⁓ you know, we live in an amazing time, but we have been frankly from my perspective for decades. And, you know, we've lived through the internet and so many of these things that are happening and I don't have to account for them all. But right now in this moment with respect to using AI to write code, I view this as a natural progression to everything that's been happening for 20 years, right? It used to be that you used to have a computer science degree. I and ⁓ I'm even starting at like the era of the PC, right? ⁓ Then ⁓ all of a sudden, ⁓ you know, all these coding boot camps spring up. They went online. You could learn this stuff online. ⁓ And as a practicing software developer, just about what like, you know, somewhere between five and 10 years ago, the truth of the matter is, ⁓ Most software developers, they're not writing algorithms from scratch and from their brain. They're often Googling around. They often end up on Stack Overflow. They copy-paste code. Exactly right. They copy-paste code right into the code base. So some of the challenges that people are bringing up even today with respect to the quality of vibe coding and things like that, we've had these challenges. And we always seem to figure it out. So what copy-pasting Stack Overflow code into... ⁓ developing a unit testing framework is the same thing here with respect to, you can't just vibe code slop, right? That is what the people say. ⁓ People who are really leveraging these tools now understand that AI, writing code and generating code for end application is just one of the many ⁓ uses of AI, using AI to write unit tests, using AI to explore the code base like Michael said, ⁓ using AI to... ⁓ rapidly iterate on specs, right? Again, ⁓ connecting the user ⁓ stories, the user intent with what exists in the code base. Product managers did not have that access ⁓ up until just recently. You'd have to come from the code base. it's exciting. There are challenges. But like with any technology, learning how to use the technology in its fullest form is... way better than dismissing it just because the one-shot attempt doesn't work. I would I equate AI to right now for co-development. It's it's you know, I have two young kids a six and eight year old. I feel like it's a six year old. It's old enough to do some things independently can give it some instructions and most of part, you know, complete them. But it'll it has you know, gets distracted by some shiny things and some new toys and it'll start diverging from what the original objective was. Sometimes even you know, contradict itself within 10 minutes of the conversation. So it's not clear, it's not like a set and forget type of tool. You can't just say, build me an app that does this and go get some coffee and come back when it's done. And so you really still as a, if it's meant to be production product and not just some throwaway experiment, absolutely. You still have to be in there and still have to be a talented developer and understand what it's doing. So you can course correct it or evaluate it to make sure that it is still consistent with your own objectives. That's a great point because I think there's a lot of this stupid scare mythology of, the developer is obsolete. It's not true. You absolutely have to have development skills and there's no way you're going to be able to software without, you know, solid. What I do feel though is to kind of pick on what you just pick up on what you said, Michael, is, you know, we have, over the course of during my own career, you know, I've leveraged offshore, you know, junior level contractors and developers because there's a place for it a value for it. but I do feel that these AI tools are going to absolutely replace that industry because you can generate equivalent quality code of these, you know, offshore, you know, low budget development, chop shops with an AI chatbot. And that industry, think is going to go away, but you still need to have the experienced manager or the architect, you know, know how to, to guide the tool and evaluate its development. But before we go to the second part, Karen, you were going to talk about those first two attempts and then those. Yeah. And Michael, I want to make sure that I answer the part of it. Like really what I was referring to is us building AI for our end customers and into the product. So we've had kind of three eras of this, right? The first one was the 2023. We should probably look at this. It's going to be important. And, you know, what is now in retrospect called the chat GPT wrapper era. Definitely untenable, useless. The list kind of goes on of adjectives that I throw on it for end applications for manufacturing, right? And LLM is trained on so many points. And actually back then in 2023, you're thinking about, know, GPT 3.5 and four was arriving. ⁓ Useless for our end application as, as, but, you know what, what it did do? was it got our wheels turning on what would eventually become the 2025 and I hope, know, kind of further extension. The 2024 era, you know, some tool sets became available. The industry started working on fine tuning these models and applying weights and using the models to, you know, so these LLMs to solve end customer problems. The problem with that was the fine-tuning, the effort was so high to get a marginal gain on what you were trying to do. So at that time, we were really, we were trying to go for a very low barrier to entry, something that we believe that the AI could get right more reliably. So rather than something like, detect this quality non-conformance on my manufactured product, we were just looking for it to be an assistant to writing queries. ⁓ against our API. Because we had a good understanding that these LLMs were going to be better at writing code than they were going to be at solving these problems. Or at least let's give it a shot. Again, the second wave, I would say, generally failed because the effort for fine tuning these models was so high relative to the game. But 2025, I would say like... There was definitely this moment in March that I think will go down as a similar moment as January 2023. GPT-5 moment? It would actually be the MCP moment, the model context protocol moment. ⁓ So this protocol already existed from Anthropic, exactly right. But it was, think, the second version of the spec, which really made it a lot more available, ⁓ established the protocol, and everybody in the industry started rallying around it. In AI, people are calling MCP, the USBC for AI, right? And I think it was just like USBC, it went through a micro and a few iterations until it got to USBC. AI has had its various kind of interface levels. And I think it was MCP protocol two or version two ⁓ that allowed the industry to rally around. got all these tools. Now you have all these agent orchestration services that ⁓ what we looked at actually just recently, the same application that I was mentioning that we tried for in 2024, what took us probably two months to get a marginal, sometimes works type of application, we were able to do in two days using MCP and these agent orchestration services. So it's phenomenal what's going on and we're in the midst of it right now. Thank you, Karen, that's awesome. I wanted to shift it now to like how you're putting AI into your two respective products because I think that's another thing. It's nice to talk about it, but how is it used concretely? And James White in the chat, and this kind of relates to, is talking about AI is a probabilistic model and we're in a very deterministic world of manufacturing. Close isn't good enough, right? If you get the wrong biom item, it's not gonna work and you're not gonna able to produce it with first Resonance. So how do we get around, how do we get to deterministic results and how are you using it in the product to obviously reduce the risk of hallucinations and increase the productivity, right? Michael, maybe you start. Yeah, I'll be too. We're using AI for our customer facing features kind of across the spectrum. But what we do, the kind of lens we look at is, it's just like a fluffy, frothy kind of feature that really has no value or is there actually need to it. And so we look, the lens we look through is, if you look back at what PLM is, PLM by definition is the aggregation of multiple data sets coming from multiple teams. You've got engineering, electrical, mechanical, procurement, manufacturing, QA, whatever it may be. In the world we live in, it's not uncommon to have these teams scattered across the world, different departments, different. cultures, different languages, different vernacular. ⁓ And so historically PLM has always had that burden of trying to transpose the data from those natural positions into a central repository before you can start processing and make value out of it. And that's put a lot of pressure on the PLM or the integration providers or whoever it may be. And so AI completely removed that dependency because one of the things that AI excels at is being able to process data in its natural format. And so now that's relaxed that burden of trying to create all these interlocking transposition algorithms to translate mechanical data into a PLM and electrical data, Chinese language and English language all into one format before it can be processed. Now we can leave it as it is. And really the burn is more on this is that digital threat. Like how do we make sure that our AI tools have access to the content? It can choose to leave it in its local repository or migrate it, but the point is it doesn't have to transposition it before it can create value. And so that's kind of the main lens we looked at in how do we add value to features. And so one of the critical areas where we do is things like the change orders, right? So change orders is where all the teams kind of come together to that watering hole and discuss what's happened. How does it impact each other? You know, whether the change order is initiated from an engineering team or manufacturing team. Often they all, good company will have all of them least have visibility into the change and they can voice their opinion about its impact on them. so again, AI does a lot of that, we'll call it, it presents the data in a silver platter to the team. Like we're not in the position right now to process it per se, because there's a lot of contexts we don't have access to, like to judge is this the right change or not, but we can't analyze the facts. What has changed? Who changed it? Is it similar to another change? What did we learn from that previous change? Can we learn those same learnings to this one? So that when the team comes in to formally review it and approve or reject it, they have a lot of that burden taken care of for them, right? That groundwork, right? Historically, change orders is a pinnacle of events and the legacy products, certainly that I've used in my career just do so much work. as an analyst, as a reviewer to understand what happened in that chain. And I would, I'm oversimplifying, but often two things would happen. I'd either just rubber stamp it and say, yeah, you I know that engineer, he's smart. I'm sure it's fine. just at which point I had no real value in the change order process, or I'd have to like budget time in my day and roll up my sleeves and go dig down and look at the original CAD models and do my own manual comparison, ask questions and do a proper review. But then that could take days or weeks. And so we took the approach of how do we combine those two where all that grunt work is done by our AI tools so that one can come in and with a minimum amount of time in an educated manner actually you will have like a time period of a rubber stamp. How about you, Karan? How did that work out for First Resonance? And the third. Yeah, right, right. And I would definitely say that we are finally in the generation of the iterations where it's going to work. I can say that confidently now. But Michael, you asked a few good questions and kind of lead up. You asked about or at least ⁓ kind of characterizing what you said mentioned we live in a deterministic world in manufacturing and A.I. is probabilistic. And that stuck out to me. ⁓ You know, this could get philosophical or esoteric, but I challenge. some of the premise, think ⁓ we want to in manufacturing live in a very deterministic world. The reality is the world and ⁓ including manufacturing operations are probabilistic. A lot of the specs that we have that we take as canonized material properties were actually recorded through a whole bunch of empirical testing in the 1960s during the Apollo era. Right? You hit this material 60,000 times. Okay, that's the hardness value. Okay. You go run that test again. this is, you know, coming from the SpaceX experience and background, I gained a huge appreciation for, ⁓ you know, going down to the physics, asking about the first principles. ⁓ And frankly, we did re-characterize a lot of material properties. ⁓ So redefining physics. So take that kind of esoteric leap forward. ⁓ What I'll say is ⁓ it's very important to build trust with your end customer. And AI is just a tool. You could say this about automation overall. One of the reasons that we have made it such an important point at First Residence to use AI inside of our business is for every single person here to gain an appreciation for the UX that's required for them to build trust and the UX that actually allows for that. So take some of these really small ⁓ UX elements that think companies like cursor or open AI or anthropic have rolled out that have developed trust with engineers. Like one example, a very simple one cursor by embedding the AI into an IDE. Again, we were talking about the text box being a profound kind of democratization and making it available to everybody. Putting AI inside of an IDE is that same. effect for engineers and developers. And counter-intuitively, ⁓ engineers don't love change, actually, right? ⁓ You know, there's kind of this stereotype that engineers are chasing the latest technology and all that, but, you know, rightfully so. ⁓ Engineers can be skeptical. They want to make sure that things are production-ready, trustworthy, high-quality, ⁓ before falling into a lot of pain of iteration. So this was, I think, an evolution. You're not copy-pasting code from chat GPT into your traditional IDE. You have the AI inside of the context of the IDE. So I bring that up to say, while I do believe that we live in a very probabilistic world, it is important for us to develop deep trust with manufacturing customers. And the way that we're approaching it at First Resonance is, ⁓ okay, let's meet in the middle. ⁓ Let us help you find what you probably missed in your initial review. of a quality nonconformance or work instruction, right? Let us surface these things that are more like anomaly detection or, did you check this? Or here's a recommendation, rather than make the claim and applying AI incorrectly in some cases, where we claim to automatically resolve an issue ticket end to end. We are far from that, not necessarily because we have all the data, but because I think it's important for us to deploy these things that say, hey, this issue might be related, or you've solved these problems 10 times in the past. Take a look here and accelerate the workflow. If it's wrong, the cost isn't that high. People aren't losing trust. But over time, the more we understand how right the AI is, the more we can actually take more bets and say, hey, we think that this is the recommended resolution to this issue. But I think, again, it's super important to not just ⁓ to leap forward and be 80 % correct. ⁓ I think it's important to say, hey, you probably missed fastening that bolt, ⁓ which is all too real of an example. In manufacturing, things are not perfect. And there are too many stories here in the past few years, past many decades, where mistakes do get made. And guess what? AI wasn't around to make those mistakes. was humans making those mistakes. How can we team up with AI to reduce those mistakes is really what we're focused on. So I don't know if you guys saw that. know, know, PTC killed LiveWorks so you don't get to hear the talks anymore. But Aora Barry, who runs AI for PTC now, he gave a speech in Salt Lake City at an event they called Fast and Furious, because I asked him, you know, about what their strategy was. And I like, he had this 3A thing for agents, the agency of an agent, where you've got an assist. I'm sorry, you have advice. So I'm just going to find the information, the contextual information, which you were talking about. ⁓ Assist, maybe you do one task. You're not allowed to do more than one. I'm gonna make sure I agree with you and then I let the AI do it. And then there's automate. Here's three or four tasks. Here's a workflow. I, in the closing part of this part of the interview was about what is your opinion of AI today? So maybe we can frame it in. Do you think we're well past the advice and we're heading towards assist and that we'll get to automate within X amount of time or do you think? There's not enough experience and there's not enough doubt in the agent today. Because I think what we're missing is an agent that says, I don't believe you. Wait a second, go back and check that again. You know what I mean? Just to the other agents. Maybe we need a committee of agents or something. So I just want to know what you guys have worked with it a lot more than I have because you've developed software. So how has your opinion changed? You started with, this is interesting. Michael, you said, well, this is cool and we got to do it because we're going to be killed otherwise. ⁓ Karen, you said, I used it already and I'd like to see it in my product. So how was your perception today and heading towards the end of 2025, what's your perception now? And the framework of, is AI already passed that advice and adding elsewhere? Go ahead, sorry, I'll shut up. No, good. Actually, I like that 3A thing. need to look up, if I can find that thought. I'll send you the, I'll put the URL in the Yeah, that's a great framework. So I had a slightly... similar but different approach. saw it, this is merely oversimplifying it, but I saw AI's really two sides, you the same point, the generation side and the analysis side. And so it's either generating content or it's analyzing content. And I personally, certainly through our experience at Duro and what we focused on, general stuff is cool. Like no doubt, right? Like I play with JetGPT with my kids, you know, like, and they get to imagine a picture and they, you know, draw me a picture of a girl on a unicorn with puppies in her hand and... in 30 seconds has got this amazing picture and the quality of those pictures have gotten incredibly better over the last year or so. Now the child doesn't have to do it exactly as five. I knew it. But it's still, in my opinion, certainly from what I see, the generous style is still nascent. I think people enjoy it and it's fun and it's certainly more mature in some areas than others, but you can immediately tell it's AI generated and there might be some hallucinations or you just have to double, triple check it. So we've really been focusing on the analysis side. How do we use AI to analyze existing data? ⁓ So you're building on top of facts. I have this data, right? I have this information, I have this build materials, I have these changes, I have these events, I have these comments. How can I learn from all the disparate information and generate better value or better intelligence for the engineering team from it, right? engineering teams and companies over the course of history, there's obviously all this, you what's referred to as tribal knowledge. You know, people know things, they don't always document it. know, what humans can get lazy. You know, on the front you see this with your customers. There's someone who's on the actual shop floor, the work instructions didn't say to tighten the bolt to, you know, 10 Torp Newtons in, but they did and they forgot to write it down. So the next person doesn't know that it has to be a 10 Torp Newtons, right? So was missing information or maybe they wrote it on a little piece of paper or their own little personal notes. So how do we use AI and the digital threat aspect of it to facilitate accessing all that information and then using AI to generate reports of value and compliments? So that's how we focus on our perspective and what we think is more mature is the analytic side. Well, the advice basically. So don't trust it to what kind of known yet. Yeah, look, I love that framework. So I'm also intrigued to take a look at that talk looking forward to getting that from you, Michael. But, you know, one word that I have used, ⁓ it also happens to start with an A, so maybe we could get this in as fourth A is augmentation. ⁓ And I really do believe that, you know, set AI aside for a moment in manufacturing and in hardware, we've always been stuck in this binary. between doing it manually and automating it. And there's a lot of stories that Elon Musk has talked about experiences at Tesla of over automating things and I experienced that at SpaceX as well. Why is this important? ⁓ You know, it is such a heavy lift to implement things like advanced robotics, AR, VR, and you know, all these promises of industry 4.0, they're very limited to like the innovation center and a lot of operations or very well-capitalized, top-tier factories where you get to do that stuff. Everything else, the 95 % of other factories that exist in the world, I you will find a very opposite view of what advanced manufacturing looks like that powers the world around us. I think AI opens up this opportunity of augmenting the human, and we're not stuck in this binary between do we do it manually or do we invest in this very high overhead automation scheme, right? ⁓ It's democratized, it's relatively cheap. ⁓ to get these answers and to take these actions. One thing I'll point out, ⁓ I actually don't believe that this framework that you described is a temporal one, where we're on some sort of phase or arc of we're in the advise phase and we'll eventually get to the automate phase. I think just as humans interact with one another, all of these are gonna be important. There will always be a time where somebody wants to do an analysis but not affect the system. And there will be a time that people are eager to take automated actions when the models are sufficiently progressed. And I think we're getting there very quickly. ⁓ One peculiarity that I noticed in every AI tool, for the most part, that we use at First Resonance, whether that's for development, for generation of prototypes, and including and kind of go to market operations. I noticed a small hint over the past few months. tool that gave this AI co-pilot capability inside of it. All of them over time provided more knobs for how to use AI. So they provide a drop-down. Do you want to ask or do you want to take action? Which model do you want to use? So very interestingly, I think in the past 12, 24 months, rather than this kind of omnimodel LLM approach where 4.0 is going to solve everything and now 5.0 is the omnimodel. ⁓ Every single one of these application layers has exposed, do you want to use OPUS 4.1 from AdTropic or do you want to use GPT 4.1? So people are getting even smarter about using models because the models have different ⁓ outcomes, just like any tool. Do you want to use a scientific calculator, a graphing calculator, or do you want to run a full FEA analysis for those hardware people on the call? AI is the same way. Do you want to affect justice function with 4.1? Do you want to refactor your own, like the whole code base ⁓ completely, you know, people say unhinged AI, right? Just go nuts. Just go slashing him ⁓ through my code base with Opus 4.1. ⁓ and do you want to ask, do you want to analyze, do you want to automate? We're going to see the same thing in manufacturing. And I think it's up to companies like Duro and First Resonance to make these things familiar, recognizable to manufacturers. What does this mean for your application? Do you want to automatically close this issue? Or do you just want to understand which issues you've come across in the past? So it's all about pattern matching. So let's move, we only got like, we're about three quarters of way through the call. I Michael's got a hard stop. I wanted to understand, now you've taken your, sold this product into the real world. And so you've, you go into, we don't have to name the companies obviously, but you go into company X, and perhaps you were able to perceive on a scale of one to five, you where they were in their digital maturity roadmap, you know. One being email and Excel and five being a fully automated agent digital twin, which nobody is doing. Basically there's no company that's at five, probably two or three companies at four, maybe. like, where do you perceive the, like when you, when the rubber hits the road, you install this. Do you get an idea of where companies are? And because you guys will each get like five minutes to respond. Like when you put in your solution, is there a ripple effect where the company says, wow, if I was ⁓ smarter about how I'd manage data and I had a chief data officer and I had data stewards to make sure everything was aligned and there was data quality, then I could get so much more out of this whole thing. Or was it just in a bubble and they did the one project and that was it? So I would like to say, how do we, to get a kind of a feel of what's happening in the real world with AI. And if it's hype or we're starting to see the industry moving in that direction. Since you got to go first, you get to go first, I would say that the problem isn't unique to AI. I think as we talked about and I brought up at the top of the call, our industry is going through its own cultural shift right now. is, I have my own theories on how it happened, but there is really a bimodal cultural effect that's happening where there's an older generation that grew up with in the 80s and 90s of those technologies. learn those and now you have a newer generation. It wasn't a continuous, you know, ingress of culture that kind of smoothed out these bumps. It really is too cultural. They're not clashing, but true cultural experiences of how they focus with technology. And so it's the latter, the, you know, more or less the younger, more progressive generation that really understands what we're doing immediately. Right. And they're already doing something like it. We're finding more and more hardware companies. go into their office and they already built something. They've prototyped it themselves because of that same factor. Like the legacy products in our product category didn't do what they want to do. Or at least it wasn't even a cultural match to what they're looking at. They're looking at these interfaces like, don't know what these are. And so they would, again, they're more empowered now and capable of building their own tools. So they often would have some type of, you know, hokey, but some type of PLM or revision management. And then they see Daryl like, this is exactly what we've been looking for. Thank you. We don't need to build ourselves anymore. And so for that, you know, that's independent of AI. It's just the more progressive software API first approach, you know, API out of the box capabilities. It's more akin to like a software infrastructure, the infrastructure as code, right? Where when you're managing your AWS environment, you can use code to configure it. You don't need to go into the You can't, don't need to go into the interface to do all the connections and tweak all the configurations. Same thing with Duro. And so that's predominantly who we attract as customers. Customers are already looking for that. understand that mentality, that approach. And so in those cases, it's more of just like, oh, thank you. We would have to sell this ourselves. We're now aware of this product and they kind of continue with their course of business. But in the other category where it is more dominated by the older generation, there is some training. and learning and you know there's like that medicine versus vitamin sales tactic you know some of them know the pains of their legacy products and what it's causing and they're looking for the medicine to resolve it some of them they think what they're using is fine they don't even I literally talk to a customer they're like what's wrong with itself and you know and then you start rattling off all these problems then like you ever have this you have this you have this like yeah we had that yeah we had that but we filled it into our business I'm like you don't have to And so that's a vitamin sales tactic, right? You need to teach them, there are better things that can help you with this ailment or prevent these things later. So anyway, I think it's a long-winded way of answering questions. I don't think it's unique AI. I think it's just unique to the more progressive staff, know, technologies that, you know, first Resonance of the bureau and many of our own other peers are providing to the industry until that first modal culture fully retires. we still have to sell to them and we still have to explain to them. And it's just, it's a different experience. You have that same ⁓ view or is your view coming from SpaceX and ⁓ the broader manufacturing than just hardware? Look, I echo everything that Michael said and you know, the analog that comes up in my mind also is cloud and not too long ago, ⁓ there was a huge allergic reaction to the cloud in manufacturing. I think the pandemic, I still do believe that the pandemic actually ⁓ flipped a switch in the collective mindsets of people just out of necessity. then people realized that, like you needed that pain in order to get on ⁓ a solid vitamin wellness journey, to use Michael's analogy. ⁓ And now people are on that. They understand that there's cost savings, things like that. I think we're still very early on in AI. ⁓ Even our most cutting edge customers, of ironically, or at least I think it's ironic, will push back on ⁓ the boundaries that we're pushing in AI. But I always go back to the kind of Henry Ford quote, right? If you ask the people what they wanted, they would have said faster horses. I think it's important for technology companies to take shots, to simplify things where necessary, ⁓ and to apply technology to solve problems that customers are dealing with day in day out because they think that's just what it is. ⁓ And you know, in our world in particular, ⁓ whether you're on that older side of the bimodal kind of generation that Michael said, or even in the middle, I will say a lot of people have gotten very used to and even have been trained on a system that is showing its cracks with respect to cost plus types of programs and waste for waste sake, frankly, and they don't even know. So what do we do about that, right? It's hard to just try to beat them over the head with it. I think it's better, more tactful, and I think this comes with some experience. Sprinkle it in ⁓ and know that you're doing good and helping make people's jobs better. Get that feedback, build that feedback loop into the product. ⁓ And just like any of these arcs that we've gone on before, We used to get huge pushback from our low code platform. We call it Ion Actions. It's basically, this is tree AI, even being able to code in the kind logical function of the factory. ⁓ So useful for things like preventing quality escapes and things like that. ⁓ We used to get pushback. we don't know how to write code. Why don't you write it for us? ⁓ What do you know about this? You should just program it in. We've been doing it this way. And then as soon as they have their first taste of the first ion action and the flexibility they got with it, we have companies that have catalogs of 50, 60, 70 ⁓ ion actions that are written into their environment. And we believe that we're going to see the same thing with AI, if not even more exponentially so, right? So it's going to take a little while. The benefit that we have, it's a double-edged sword, is we get to see what's happening at the leading edge, right? How software developers are using AI and what consumer applications are doing with AI. And then we get to work backwards into what can this look like in manufacturing a year, two years, five years from now. And I think it's important to continuously lean forward because as Michael said, if you don't do it, your competitors will. And I think for us in particular, if we don't do it in manufacturing, there's a risk that no one will. And we revert the entire system back to ⁓ Back to, know, I don't ⁓ know. I don't like to go visit those places, but it is something I ⁓ hold very, ⁓ is like very important to what drives me every day and why we're taking these bets on technology taking us forward, not backwards. So a slow ripple for you. Yeah, slow at first and then very quickly is what we've seen in previous iterations of this. We're still very early on the AI side at first resonance. But ⁓ I'm very much looking forward to reporting back in about six to 12 months time. All right, well, let's do that. Let's have another one of these calls together. This was great. Really appreciate your time. I learned a lot. That was pretty really cool. I hope you guys enjoyed it. Awesome. All right. Well, let's do it again. you know, good luck with the fourth quarter because it's starting up pretty soon here. ⁓ great. It's awesome meeting you, Karan, because we didn't know each other. Michael, I look forward to seeing that demo someday. I'll take it easy. It was awesome. I look forward to doing this again in a couple of months maybe. Awesome. Thanks a lot for saying that. Likewise. Thanks both of you guys. Great discussion. you. Take care. Take care.