The 80/20 Rule for AI in Manufacturing — with Dirac and LimitlessCNC

Episode Summary

The episode titled "The 80/20 Rule for AI in Manufacturing — with Dirac and LimitlessCNC" delves into the evolving role of artificial intelligence (AI) within manufacturing and engineering processes. Hosted by Michael Finaccaro, the conversation features David Priev from Limitless CNC, an AI startup focused on developing automated CAM programming technology for CNC applications, and Filip Aronstein from Dirac, which has created an automated work instruction platform designed to enhance efficiency in manufacturing environments.

The episode highlights two key insights. Firstly, both guests emphasize that AI tools are likely to act as force multipliers rather than full replacements for human engineers. They suggest that these technologies will handle 80% of the routine tasks while humans manage the remaining 20%, ensuring a balance between automation and human oversight. Secondly, they discuss how their respective companies leverage AI to address complex manufacturing challenges, particularly in areas where tribal knowledge and context are crucial.

For PLM and engineering professionals, the key takeaway is that AI should be viewed as an augmentative tool rather than a disruptive one. By integrating these technologies into existing workflows, engineers can focus on more creative and strategic tasks while leaving routine operations to AI-driven solutions. This approach not only enhances productivity but also ensures that critical human expertise remains integral to the manufacturing process.

---

Full Transcript

Michael Finocchiaro

Okay, we're live. ⁓ Welcome to the 13th edition of the AI Across the Product Lifecycle podcast. This is Michael Finaccaro, and I'm joined by David Priev of Limitless CNC and by Filip Aronstein of Dirac Do you guys want to introduce yourselves?

David Priev

Yeah.

Fil Aronshtein (Dirac)

Sure.

Michael Finocchiaro

Go ahead, David.

David Priev

So pleasure being here, Michael. Thanks for the invite. I'm David. I'm the CEO and co-founder of Limitless CNC. We are an AI startup based in Israel, Tel Aviv, and we're developing AI technology for CAM programming for CNC applications.

Fil Aronshtein (Dirac)

Thanks. that.

Michael Finocchiaro

Awesome. How about you Phil?

Fil Aronshtein (Dirac)

I'm Fil I'm the co-founder and CEO of Dirac. We've built what we call the first and only automated work instruction platform. Originally an electrical engineer and roboticist by background, started the company when I got my start over at Northrop Grumman or started the company after I kind of saw how unfortunate and archaic a lot of the infrastructure was in the manufacturing side of things. And so I wanted to do something about it. So that's us.

Michael Finocchiaro

⁓ So, you know, our talk tends to be ⁓ about AI and how it's changing the role of mechanical engineers. And I think all of us here are mechanical engineers, if I'm not mistaken. So, you know, if we look back at two, three years, we went from AI as machine learning and some LISPs up for decades before to, my God, it's part of my daily life and I can't live without it anymore. ⁓

Fil Aronshtein (Dirac)

Good. Let's get this.

Michael Finocchiaro

However, there's a lot of skepticism on how applicable is it in the engineering world? Is it just hallucinations? You know, how do you put guardrails on it? So I'm just wondering, like, if you look back to, you know, 2022, 2023, Phil, were you like really bearish on the potential of AI to transform manufacturing and, you know, engineering, or were you a bit skeptical or you're sort of wait and see?

Fil Aronshtein (Dirac)

So generally speaking, think non-deterministic solutions are never going to be 100 % replacing people. ⁓ think generally, and I felt this way in 2022, I feel this way still, I think AI tools are going to be force multipliers for folks. I think they're going to generally be 80-20 tools, where you'll have an AI system do 80 % of the grunt work for you, and then you'll have a human in the loop. you know, doing the final little bits and pieces of it. Because in manufacturing, lot of the thing is about a lot of these models, they are trained on the internet. They're trained on data sources that are out there. a lot of the, yeah, exactly. The thing is it's like in manufacturing, a lot of what you need to do a excellent job is tribal knowledge. It's context that like doesn't exist in a model. And so ⁓ that's, I mean, that's how we build our product. That's how we think about our product.

Michael Finocchiaro

We're on Reddit basically, right? On Reddit.

Fil Aronshtein (Dirac)

Folks on our software team use AI. It's very, very 80-20. I think it's going to be that way.

Michael Finocchiaro

Thanks Phil. How about you David?

David Priev

So I have to say that ⁓ I believe that we're just in the beginning of seeing what we can do with AI and the engineering world and specifically with like the 3D world, CAD, physics. ⁓ know, everyone now thinks that LLM drives everything, right? Like I had so many discussions with customers, being confident that we're probably running some LLM under the hood, right? AI is LLM, right? ⁓ But the fact is that I believe LLM is just the first wave of AI bringing the revolution to humanity. And we're going to see tons of new stuff in the next few years, especially on the physical side.

Fil Aronshtein (Dirac)

Thank

Michael Finocchiaro

So, Filip, you were a little more skeptical and David, you were a little bit more bullish, I guess we could say, right? ⁓ So, another thing that ⁓ is in the news a lot is VibeCoding, And Novabool has this insane market evaluation and ⁓ market valuation, sorry. ⁓ But at the same time, mean, I don't think that anybody is developing software today without using, you know, Cursor or windsurf or whatever. How, how is, how are your teams, David, using AI and is it like the scaffolding is it for creating PRDs or is it actually for coding? And is there a difference between like the junior programmers or the senior ones? Like I talked to one startup and he said, Oh, I don't give the tools to the junior programs. I want to learn how to program, but I give it to the senior guys so they can go faster and be and help them. So I thought that was. an interesting way of doing it. I'm very curious to see, you know, from your perspective, David and then Filip.

David Priev

So first of all, I'm a big believer in those code tools, but I think they're just tools, right? Like they're not going to replace you and I don't believe that you should do like vibe coding the entire day. Now, what I see lately, know, like every month or so you see like the new model is like performing much better by Gemini and then it's Claude and then...

Michael Finocchiaro

It's bad for you, right?

David Priev

So what we try to do here at Limitless is, first of all, allow everyone to have access to all the latest models and just let them try it themselves. It's amazing to see the progress that those companies ⁓ are making with those tools. And lastly, I would say that I see it very relevant for now, like for classic SaaS applications where you need some front-end components, straightforward backend ⁓ components, but when it comes to deep algorithmic challenges, especially with niche, understanding how BRap representation, how it works, how CAD are running, it's just scratching the surface when you try to work with those AI tools. So I think the core of what we do at Limitas is still very human-driven, I would say.

Fil Aronshtein (Dirac)

you Thank

Michael Finocchiaro

Okay. And yet you're using AI as a way of accelerating and so forth. How about you, Filip? Same thing or totally different?

Fil Aronshtein (Dirac)

would say very similar. would say everybody on our team uses AI tools, they use them aggressively, use cursor, use all that stuff. I am very much like a proponent of this. I say, here's the company credit card, go rip as many credits as you want. We need to move fast. Velocity is our advantage. you can put money in and get productivity out, like go do it. I'm very, very much in favor of it. But like, as David says, if you're building like a neat little like CRUD app that is just like a database and a cool little UI and it's like typical like B2B SaaS, it's easier to do stuff like that. But when you are working with like really complex like physics simulations like us or any sort of computational geometry infrastructure, it is

Michael Finocchiaro

Go for it. Yeah.

Fil Aronshtein (Dirac)

really, really tough. like, you know, the LLMs like don't really work super well with stuff like that. you know, we were sort of fortunate in that we built, I think, the majority of our like really complex infrastructure. We built it before like all the AI stuff got really good. And I think now that a lot of the LLM stuff and all the like vibe coding stuff has gotten really good. I think a lot of software engineers are like, it's gotten very easy to build like very unsophisticated stuff. And so the amount of like effort and dopamine that a software engineer is like willing to or gets out of like building something like that, like there's, there's almost this like a lower attention span for a software engineer on average now, because they are so used to being able to get like really easy front-end, like not super complex infrastructure. And so I actually think there's this like interesting moat that companies like, you know, Dirac and companies like Lumen and CNC end up having.

Michael Finocchiaro

Yeah

Fil Aronshtein (Dirac)

because we built some really, really ⁓ complex stuff that is not easy to vibe code. nobody is, it's way tougher now for a software engineer to ⁓ look at something like what we have and say, yeah, I'm gonna vibe code that. And then the second they can't, they're like, all right, onto the next thing, I guess. And so, yeah, I think the context windows that a lot of these things have also not super great. And so like, you know,

Michael Finocchiaro

Yeah.

Fil Aronshtein (Dirac)

You very, very quickly, if you're like using Vercel v0 to try to like vibe code something or any of the other coding things, um, you very quickly like reach the limits of like the complexity that you can vibe code. so I like, use them for, we use vibe coding for stuff like writing a PRD or writing features. would say like our, um, our, guy who runs product on our team, Jared is like an exceptional Figma guy and like is a manufacturing engineer by background. So like deeply understands the problem set. is a software engineer by background as well. ⁓ And so he can like use AI tools effectively. But like that's kind of my philosophy on like if and when and how to use them. If you do not know, if you can't understand the code that the LLM is writing for you, should not be using it. And so like I compare it to like for the guys on our team, I compare it to being sort of like a software engineers are now like sorcerers. They're like wizards. that can cast ⁓ imps and create this imp to go do my bidding and do an agent to go do this mindless task. If you're a level 100 wizard and you could just spin up a million imps to go do a whole bunch of different tasks for you, awesome, great. But if you could not have written that code yourself, you were going to enter a whole world of pain when you actually tried to integrate all of these different vibe coded modules that you tried to

Michael Finocchiaro

You

Fil Aronshtein (Dirac)

So, yeah, exactly.

Michael Finocchiaro

or build a test harness based on it even worse. Yeah, actually I just, through a friend of mine, discovered anti-gravity and that's been kind of insane because you can actually create agents for every single coding thing you want to do. It's just, it's impressive. But it's cool too to say that, you know, the fundamentals of being an engineer, fundamentals of programming are still critical because as you said, you have to be able to read the code. If you can't read it, And when it screws up, how the heck are you going to debug it, right? I guess the one awesome thing about it is how good it is at documenting itself, though, which is what most programmers are really bad at. What the hell was that loop? That's awesome. So OK, both of you guys are pretty aggressive in using AI and development. How about when I talk to Mauro of Leo AI, another Israeli startup, Mauro actually developed his own LLM, right? Or an LMM, he calls it Large Mechanical Model. I had him on one of my first podcasts. ⁓ How are you guys using the LLMs? Are you thinking of building your own model? You know, in your case, David would be more on ⁓ machining and CNC and previous tool paths and so forth. Or maybe Phil would be more on the work instructions on how did I build this thing the last time or last 100 times or five years ago? I mean, are you guys thinking of that or is it more, you're using more rag to bring in that information, you know, as you're developing the calculation, how are you integrating that LLM power into the way you guys work? Either of you can pick that one up first, I don't mind.

David Priev

So at the beginning when we started Limitless, one of my co-founders is our chief scientist. He has a PhD in computer science and he's an expert in 3D and AI. He took a brave decision ⁓ about LLMs that it's not going to be part of our core technology. ⁓ From the main reason that LLMs are great when it comes to text, and ⁓ context. You have a context and then you want to understand this context and come up with ⁓ an answer that is relevant for this contact window. But when you look on CNC machining, basically, first of all, the domain is not text, it's geometry and it's physics. And more than that, it's not that of a matter.

Fil Aronshtein (Dirac)

Okay. So, thank you.

Michael Finocchiaro

Always, yep.

David Priev

what was the previous steps that brought you to the current material shape, right? It's all about what's going to be the next move. So it's more similar, I would say, to like playing chess, where every board ⁓ snapshot is like, it's like a new game, right? And now you need to think, okay, what should be the next move? Having said that, like we see a lot of value incorporating LLM on the product level, and I'll explain. So I think LLM,

Michael Finocchiaro

Hmm.

David Priev

⁓ and, and in general, all those chat GPT types of tools are doing great by educating users to interact with software by, by, like writing to it, by communicating with it. And I believe LLMs can be great to replace like bunch of complicated UI workflows and buttons and, and, and all those stuff that you usually need to use in the current software, especially for CAD. So let's take, for example, like all the.

Fil Aronshtein (Dirac)

Thank you. So,

David Priev

huge toolbars in CAD softwares, right? You have like tons of menu wizards, tons of clicks. Now go figure it out what you're doing. What if you had just one search button where you just text what you want to do and then it shows you, okay, this is what you want to do, right? And then let's do this. So I believe that LLM has great advantage in simplifying the user experience on the product level. And this is where we will incorporate that for sure.

Fil Aronshtein (Dirac)

Mm-hmm.

Michael Finocchiaro

How about you Phil?

Fil Aronshtein (Dirac)

Very, very similar mentality. Like our core infrastructure, our core, uh, like utility is deterministic. Um, it's all physics simulation based. It's a lot of computational geometry and really it's a lot of mechanically informed heuristics. And so like the core skeleton, the core utility, the product is itself deterministic. Like we are building at its core, really like a new feature, a new product, a like building our own category of thing that like today does not exist. Like the first animated, automated. like model-based work instruction that like thinks about integrating the digital thread into the work instruction and vice versa and aggregating tribal knowledge and putting that into the digital thread. And so like we thought like, okay, let's lay down that foundation infrastructurally. And then like in the way that we were building the software, let's keep in mind that we're going to want to have some like probabilistic solutions. We're to want to layer. And so things like classification of components, things like having our own ⁓ understanding of what components are and how do they physically interact with one another, having this sort of like ⁓ mechanical systems themselves, like assembly itself has a language to it. And so the way in which we are over time cataloging and codifying the theory of mechanical engineering lends itself very well to the architecture of the transformer, ⁓ although like you know, not necessarily us throwing like text generation at like different things. Although very obviously like generate semantically relevant text for like a work instruction is useful, but like frankly not even as useful as you know, people might think. a lot of the, like genuinely it's like ⁓ people say like a picture is worth a thousand words. We like to say an animation is worth a million. And so if you can actually in this 3D animated environment show an assembly coming together, you don't really need words.

Michael Finocchiaro

Okay.

Fil Aronshtein (Dirac)

Like if you look at the work instructions that like a company like SpaceX has, one of the most advanced manufacturing companies in the world, they don't really have words in the work instructions. It's pictures and icons. It's like their whole thing. ⁓ That's exactly it. You really want work instructions to be like as simple and straightforward to ⁓ understand and read as humanly possible. It should have like a very, very low ⁓ management index in that.

Michael Finocchiaro

Just like IKEA, ⁓

Fil Aronshtein (Dirac)

if a manager were to get on the production line and start doing the assembly themselves, they wouldn't actually mess up or hurt themselves. And so you can't just rely on operator tribal knowledge. so ⁓ words are often not even like the best thing to describe stuff like that. a lot of like our core mentality is understanding geometry, understanding physics, understanding how components interact with one another and being super modular and automated.

Michael Finocchiaro

Cool. Yeah, the... The idea of the instructions being as easy as Lego or IKEA. That's pretty cool. Because a lot of the work instructions tend to be kind of hard to understand. ⁓ So in the actual product and in Dirac and in Limitless, how is the... Well, I think you've already almost answered, David, but the user comes in contact with AI how? I suppose that you were saying that the user gets it from the user interface from the beginning. Is there any AI under the covers? Is there a little bit of work or maybe it's ML and not really AI in order to ⁓ learn from previous experiences and predict what the next tool path will be? Because you're saying that's whole thing is the next step.

David Priev

So ⁓ actually there are two deep technologies that we are incorporating into our ⁓ engine. The first would be what we call our feature recognition engine. And everyone that are in the CAD domain know what I'm talking about, right? You eventually need to understand what you see. ⁓ Exactly. The problem is that the current industry standard for feature recognition is mainly focused on

Michael Finocchiaro

Chambers and Phillips and.

David Priev

CAD features, like as you mentioned, chamfer, radiuses, slots. But when you talk about machining, ⁓ machinists are looking absolutely different on geometry. They put in their mind what should be the machining operation to apply on those group of surfaces. So basically what we're developing is a neural network that can receive CAD in its native format. And this is important to emphasize.

Fil Aronshtein (Dirac)

Mm.

Michael Finocchiaro

Hmm.

David Priev

and can label the entire CAD surfaces into the right groups and into the right labels of the machining operation that should be applied on those areas. So we call it, like, this is the perception capability of our AI. And we train it on real-cam data. So this is another thing, like, how do you train your network on this? And second part of the technology is what we call our physical AI engine, which is ⁓ in a very simple way. It's a decision tree algorithm that try to pick up the best next move to do, the best next step to do out of different options using a sophisticated scoring function that is mainly based on physical estimation of different phenomenons in the process like ⁓

Fil Aronshtein (Dirac)

Thank Thanks.

David Priev

the chatter or the natural frequencies that the tool can get to based on the geometry of the tool, the deformation of the tool, the deformation of the working-pest geometry based on the cutting force, like all those different physical phenomena that great machinists are simulating in their head with amazing intuition, by the way.

Fil Aronshtein (Dirac)

you

Michael Finocchiaro

Yeah.

David Priev

This is exactly what we try to incorporate into the decision-making engine. So I would definitely call the two of them deep technologies. They're much more advanced than just straightforward statistical approach. But it's not LLM as a lot of people want today ⁓ to see. LLMs are like everywhere. So people think that that solves everything, but it's far from that.

Michael Finocchiaro

But you were saying that you already have that chat interface for the front end from the user perspective too?

David Priev

Yeah, and in general the way that the user interacts with our product. ⁓ So it's a copilot that is like ⁓ running alongside the CAM software, so it's not replacing the CAM software.

Fil Aronshtein (Dirac)

So, yeah. Thank you.

Michael Finocchiaro

so it's like an add-on to Mastercam or Gibbscam or whatever?

David Priev

Exactly. On the backend, it's integrated via API to the cam session and can execute the recommendation directly to the cam environment as if you did that. ⁓ And on the UI level, so it shows the user what is the next recommendation. And the user can like give feedback, change it, accept it if he accepts the recommendation, which is generating it straight inside their cam environment. And then we'll move on to the next suggestion. So this is the current ⁓ user flow.

Michael Finocchiaro

So it's sort of one of Filip's imps sitting on his shoulder, looking over his shoulder saying, do that, do that. So what about your imps in Dirac then? ⁓

David Priev

Hahaha

Fil Aronshtein (Dirac)

Thank you.

David Priev

Exactly.

Fil Aronshtein (Dirac)

So, broadly the way that we think about like integrating LLMs into like feature sets, ⁓ kind of like what I was talking about with ⁓ like semantically relevant text, ⁓ stuff around how we ⁓ like, the work instruction itself serves as like a really interesting place where an enormous amount of like context and tribal knowledge like finds itself and the way that we've like built features and functionality in the software today is very structured. And so like we can give you like in the same way that you ⁓ like might want to like query Google for like context information. We were building effectively this like context aware production planning platform that you you can actually query and like index and say, Hey, ⁓ when was the last time I used Loctite 242 and like what parts did I use it for? And who like, you know, when those work instructions were approved, who was actually at my company and who signed off on that? And, you know, what happened with that part? did that, you know, Loctite 242 find its way into like the following rev of that work instruction? Or like, that replace it? So a lot of that context lives inside of the heads of folks who've been on the shop floor for like 20, 30 years. And we are finding that like the work instruction is this really excellent place for us to just aggregate all that context. and surface it to folks where they might find it useful and even, you know, bringing it earlier into the design phrase. That's a lot of how we...

Michael Finocchiaro

Interesting. It's also interesting too, you guys are really complementing each other, right? On one hand, David's doing that for the machinist and you're doing it more for the guy in the assembly line, right? That's interesting. So, what about the, know, where do you sit today? mean, you already talked a little bit about your vision, but now we're, what we saw this year in 2025 was Anthropics MCP, which was another, a big earthquake in the way that we thought about AI becoming autonomous and having agents, talking to agents and executing entire workflows. What is your feel about that? Where are you going to go with the agentic thing and where do you think is coming next? I don't know if any of us who were really expecting the agentic thing to go that viral so that quickly.

Fil Aronshtein (Dirac)

I mean, I expect us fully to have an agent that understands how to use Build OS and it could just rip work instructions for you. ⁓ That's kind of the idea. it's funny, if we were thinking about mechanical systems and mechanical design like 20, 30 years ago before CAD was very, very sophisticated, You would have to have built CAD software first, and then you throw an agent at it to like make CAD models. And we're kind of in like a similar place where they're just like, is not a robust enough work instruction platform to do a sophisticated stuff as what we're allowing folks to do. And so it's like phase one is you build the really sophisticated platform for the thing, and then you can like throw an agent at it and stuff. That'll, that'll probably be something we work on in 2026, not like a Super I mean people are asking us to do it not a very hard thing for us to do and build it's just you know On the roadmap one of those things

Michael Finocchiaro

And you, David?

David Priev

So we're actually gonna deep dive into the physical integration, I would say, with the machine floor. idea now, like currently we help to program the part, right? But you know, like how the process looks like today, you program a part, you run it to the machine, right? And then usually it never goes well. You take it to the quality inspection, you run the CMM. By the way, you also need to program a CMM program. program that has a probe that runs on the geometry and gets you the real dimensions and the offsets from what you need to achieve. So you do that and you look at the report and you see, okay, wait, I'm missing here. I'm out of the tolerance. I need to change something on the setup. I need to change something on the program. This is what usually happens. We call it the loop, right? ⁓ And currently this loop is not automated. It involves zero

Michael Finocchiaro

Right.

David Priev

technology and AI, especially something that helps to take better decisions. And this is where the expertise needed. So this is what we try to ⁓ achieve at Limitless. First, helping with the programming. Then in the future, we're going to introduce our machine connectivity. So we're going to connect to the machine controller, fetching all the data from the real toolpaths, cycle time, all the machine sensors, ⁓ and all this data will go back to the agent. And more than that, we're to also incorporate the dimensional reports from the quality inspection, putting that back to the agent as well. Now imagine you open the projects on the second time and the agent already like figured out for you what you, what should you change in the process in order to improve it. And that's going to save so much time. And, and it's going to like lower the.

Michael Finocchiaro

Yeah, of course.

David Priev

high bar of expertise required today in order to solve those complex parts and I'm talking about parts like you know in aerospace, defense, medical stuff like complex geometry, tie tolerances. This loop can take

Michael Finocchiaro

There's stuff you're running around Tel Aviv to find in this story I wrote about Limitless. Is there going to be a Limitless CMM then? You're going to do a CMM module too? That's what it sounded like.

David Priev

This is also part of the process. By the way, it's a very similar technology. It's still like a CNC machine. It just runs the probe and not cutting the material.

Michael Finocchiaro

Yeah. How about you, Ditfil?

Fil Aronshtein (Dirac)

I mean, we're not going to get into CMM ourselves, you know, but a lot of how we think about like what we're going to be building in like the future is us getting into like automated factory layout planning, automated maintenance and repair instructions, know, automated inspection and quality routings. I mean, we're actually getting into DFM right now. A lot of how we actually dub ourselves the assembly company. A lot of people care about how to make parts. We don't. We care about how to put them together. And so that's

Michael Finocchiaro

Thank Very cool. Yeah.

Fil Aronshtein (Dirac)

A lot what we focus on is just how, ⁓ you know, how are we doing production planning in a facility? How can you take context from one person in one workflow and bring it into another? Like we think every piece of data, every piece of information in a manufacturing facility is one node in a very deeply interconnected graph. And so when we look at the work instruction, it's sort of like a bundle of these nodes and we just draw a rectangle around it and we call that a work instruction. ⁓ But, you know, the industrial engineer doing factory layout planning has this like top-down macro perspective of work instructions and how they interact with each other through space time and that is their perspective which you look through the lens of like factory layout planning and you know doing a bunch of different things that they end up having to do and so that's kind of like how we intend it, you know build stuff out ⁓

Michael Finocchiaro

So it sounds like you could almost, well, you could even do the service notes in a sense.

Fil Aronshtein (Dirac)

Yeah, yeah, definitely.

Michael Finocchiaro

That's really cool. thought if I could ask, mean, you guys are both doing ⁓ stuff that arguably the big three should already be doing, right? mean, already Siemens and DS have invested billions in manufacturing and Inovia has work instructions and they bought a Prezo. I mean, where's the gap? How are you guys able to compete? Because you think the customer like, no, but the DS guy sold me the thing and

Fil Aronshtein (Dirac)

What is this?

Michael Finocchiaro

I mean, it's pretty awesome that you're able to do it. And I'm just wondering what's the secret sauce, because probably other startups will listen to this and be like, but no, I'm not going to, or maybe someone listens to podcast and they'll be like, no, I'm not going to do that because I'm going to be killed by the big guys before I even get out of, get my idea out. Cause I find you guys super courageous, you know.

Fil Aronshtein (Dirac)

Yeah. It's a really valid question. ⁓ So for context, we have a really, really great partnership with Siemens. They've brought us into customers. We're going after customers together. really, really tight. We have an integration with Team Center. We have really, really good relationship with these guys. And I've often asked myself the same question, right? Why have the big guys not built something like this? ⁓ And one can make a whole bunch of different arguments. You know, the canonical classic startup argument that a founder makes is always like, the big guys don't build anything. You know, it's, you know, why won't Google build this? It's like, oh, Google doesn't build it. You know, it's, I find those arguments to be like somewhat compelling maybe, but like, I think it's really just like the technology and the infrastructure that they have built over the past 20, 30, 40 years is pretty. sophisticated, but what they are exceptionally good at is not necessarily building, it's selling. Like their power is in their distribution, their power is in their relationships with their customers. The number one thing they have is customer trust and customer loyalty because the thing that a massive aerospace manufacturer cares about almost more than technology and capabilities ⁓ is the company that I'm working with, is this software, is this technology going to be around in five, 10 years? And so, know, big companies have to work with big companies. It is very, very tough for a startup to earn like the trust of a prime or of a massive automotive manufacturer that they'll be around in three to five years. You know, I would say that is one of the key things you have to do in building a software selling to those companies, the big guys, you the enterprises, ⁓ and especially in a way in which you like might be adjacent to or might have to integrate with. the big three, you're going to have to make a really compelling argument to the big guys that you're going to be around. ⁓ so I think we've been able, we're very fortunate we've been able to do that. We've been able to show that we have this 1,000 times better piece of software that does not exist for anybody today. It's a real, real massive pain point for people. And so we've been very fortunate that we could solve that problem so aggressively for folks and also be able to integrate with the existing ecosystem. I think both are very, very key.

Michael Finocchiaro

It's really stickiness. It's really for you one of the most important things. ⁓

Fil Aronshtein (Dirac)

Yeah. I mean, our whole perspective is like what CAD did for mechanical engineers. We have built this like 21st century upgrade to the workflow for manufacturing engineers, something that like does not exist at all today. And so if you can make that type of argument, then you you, you, you, you could stand a chance.

Michael Finocchiaro

David, what's your experience?

David Priev

So ⁓ first I would say.

Michael Finocchiaro

Of course, you're in a more fragmented world though, right? Because the cam world is relatively fragmented. There's no huge, well, Sandvik, but Sandvik has two different, well, actually five different cam packages, right? So sorry, it's not exactly the same.

David Priev

So yeah, it is a fragmented market, but when you look at the enterprise landscape, you will still see like the big three, four players dominating in Anex, 3D experience in KTIA, PTC, by the way, as well, ⁓ Mastercam.

Michael Finocchiaro

NX Cam, yeah. And then Autodesk, mean, Fusion's got some great manufacturing stuff.

David Priev

Yeah, although fusions are more looking into like small, medium market businesses. ⁓ Anyway, what I was trying, what I want to say is that, like, let's say I will be very unconfident trying to build the next CAD or CAM platform to compete with what exists today because those tools are like, they went through so much, Like they're so... solid and it's so hard to transfer a client from one data type to a new data type. So it's almost mission impossible, right? ⁓ But if you think about a solution that is like a new layer, right? You don't need to change anything in your stack. ⁓ And this new layer actually brings a lot of value without changing anything and it's integrated into your workflow. So there you go. You have like a new opportunity here. the market. This is what we're trying to do with this agentic layer of a copilot that runs on top of cam. ⁓ And regarding the competition, so yeah, for sure they have teams that working on this, there's two key things that differentiate, in my opinion, like startups like us and the large players. First, ⁓ they need to maintain like a legacy solution, right? So how much of their energy actually goes to like changing

Michael Finocchiaro

Technical debt.

David Priev

core capabilities, like reinventing the wheel there. Not a lot of energy goes there, right? ⁓ Secondly, ⁓ I believe that young startups with a compelling vision and a good funding backing can recruit amazing talent that usually will not think or consider to go for those large players. as an AI career, right? And this critical talent, especially when it comes to AI, can be a game changer.

Michael Finocchiaro

Yeah. Absolutely. ⁓ Actually, I just got a question on the chat from Matt Ferreira, who says, do scan programs usually compute the cutting forces and temperatures on the tool in part? Does limitless calculate that? And are you able to detect things like force spikes ⁓ when cutting inside corners? Good question, right?

Fil Aronshtein (Dirac)

Okay, I think that's it I'm going to start my program. We'll see you next time.

David Priev

That's a very technical question and I like it. So first of all, ⁓ 99 % of CAM softwares do not calculate those type of stuff. CAM and CAD, by the way, CAM brought into the world out of CAD, right? You had CAD at first and then we say, okay, we can do...

Michael Finocchiaro

Actually, could argue it goes the other way because ⁓ the guy that invented Asus, he actually started doing a manufacturing software before Asus became the CAD software. you're right. is a big interplay.

David Priev

That's interesting. Like if you look at Siemens, Autodesk, they saw those guys started from CAD and then they brought the CAD ADA into CAM. Before that you had G-code, right? ⁓ So all of those softwares are basically very geometry oriented, kinematics oriented, but has nothing to do with physics exactly. When you want to incorporate physics, you either need a third party solution like VeriCut that can simulate the toolpath and calculate the forces.

Michael Finocchiaro

Good. And that's a physics oriented. Yeah.

David Priev

and the process and then it can calculate the parameters that were raised. ⁓ Or you need to go like deep into like finite elements, start like, you know, building and modeling the process on your own. Like there's no out of the box features in Camden doing this today. And this is where the value and the opportunity is. Like this is what's happening inside the machinist mind all day long when they program, like they think about those stuff. So we want to bring that into the AI.

Michael Finocchiaro

And then, and so your answer to that question is yes, Limitless is aiming to do that.

David Priev

is taking that into consideration in the physical engine, but we're not showing the user all those different calculations.

Michael Finocchiaro

Right. All right. I actually have a friend of mine that's a NX cam guy and he was telling me about how, how much physics is in there between the, the velocity, actually the acceleration deceleration, how you get rid of the excess material. I mean, it's just, and he said the math is just more or less insane. ⁓ about you, Filip, ⁓ we've been talking too much about cam and not enough about work.

Fil Aronshtein (Dirac)

I mean, hey, I mean, I love hearing about Cam. It's kind of funny. You would imagine that people who know about CAD would like happen to know a lot about Cam, but like actually not out of time. I have friends who I am daily. have friends who I literally have a friend named Cam who like it's incredible company going to plug in for a second and company called Rangeview.

Michael Finocchiaro

Hahaha

Fil Aronshtein (Dirac)

They do casting, incredible company. And I would say, I hear a lot about Cam Tools. I hear a lot about how they're advancing. ⁓ I generally think that the way in which software on the manufacturing side and the design side has developed has become almost a little too generalized. I like that there is this more, like, Michael, know Brad and Tom. Like I'm a very, very big fan of NTOP and of Brad, ⁓ but like they're very, very focused on parts. ⁓ And I think literally speaking, like CAD software has become a little bit too generalized in that when you were building out like an assembly or a system, you can design a part and you can also design an assembly at the same time. And it's almost like you're not allowing yourself to pin down enough, ⁓ pin down enough variables to be constants. And so you, you are building systems in like too flexible of a way.

Michael Finocchiaro

Hmm.

Fil Aronshtein (Dirac)

Like it's almost like I'm totally taking us on a tangent here, but you know, my, lot of my soapbox is like in order to manufacture really quickly and effectively, you need to pin down more variables. You like cannot necessarily allow designers and engineers to have as much flexibility and as much creativity as like they might want to and currently have in CAD software because in CAD you have this sort of like

Michael Finocchiaro

You

Fil Aronshtein (Dirac)

know, limitless 3D environment and you can build like this platonic ideal of a system. And by the time it like gets to a manufacturing engineer and their jobs to be like, okay, cool, got this CAD file. How do we actually make this thing? ⁓ You know, they now have to like rework the design a ton to be ⁓ designed for manufacturability. And a lot of like how we think about stuff, a lot of how I like hope we start pushing more and more is adding more like manufacturability, like real world constraints. bring that earlier into the design process such that CAD is a little bit more guardrailed around the world and design is a little bit more guardrailed. that's kind how I mean, the CAM conversation was making me think a little

Michael Finocchiaro

Right. Right. That reminds me when I interviewed ⁓ Sohrab Haggigad of Hestis and he was talking about how ⁓ the problem with 3D printing is that the generative design is you get these organic things that are not manufacturable, right? And if you don't have a 3D printer and you really have to do it in a milling machine, you have to redesign the whole thing. So his whole program was, his startup was that now I'll give you the real geometry that's manufacturable. It also reminds me, I was invited, I missed out on the conference in Austin because I was invited to speak at a conference in Scandinavia in Copenhagen and they sort of asked me to come in and they created a thing called ⁓ RD8 and I thought, Radi8 and they said, no, no, Rugged Design 8. I'm like, okay. But Rugged Design, apparently, is a whole thing that they do in Scandinavia and it's you look at a design and you add the constraints, you look at the constraints and you find like the thing. It's basically what you were asking for, Phil. It's pretty interesting. I'll put you in touch with him. ⁓ So let's talk a little bit more about when you take the thing into, you take it out of the little ⁓ Paradise Garden or the Plato's cave and you've now put Dirac and Limitless in the real world. ⁓ two questions. First question is, I look at digital maturity on a spectrum of like one to five. One is like, I'm still doing email and using Excel and five is like autonomous, agentic, ⁓ adaptive digital twins. And basically nobody's at five, right? Barely anybody's at four, maybe a handful are at three or think they are. ⁓ So, it is in your experience, where, do your customers sit? And maybe it's an industry thing. Maybe some industries are more advanced than the other. Like for instance, if I just throw one in, into, I talked to a company called Eximpro. Peter is very, active on LinkedIn. You've probably seen.

Fil Aronshtein (Dirac)

It

Michael Finocchiaro

Peter Spohs, he told me about the mining industry. The mining industry came from the Stone Age in terms of computer stuff. In the last 10 years, they're now one of the most advanced. They have these awesome autonomous vehicles. They're going up and down the mine. have the most awesome battery technology on the planet because an incline like this with carrying ore on it, the battery is dead within three seconds. It's insane. They went from nowhere to almost a four or five, almost all five big mining companies.

Fil Aronshtein (Dirac)

Yeah.

Michael Finocchiaro

You guys aren't in mining, you guys are doing other stuff. What's your opinion on that?

Fil Aronshtein (Dirac)

I've never run into a five for sure. I don't think fives really existed in manufacturing yet. I've not really a four yet either. I haven't really seen any, maybe like one four and I was sort of blown away. I'd say the majority of our customers fall into one to three. I would say it's probably a little bit of a bell curve. Like sort of, mostly in two, a handful in one, a handful in three.

Michael Finocchiaro

yet.

Fil Aronshtein (Dirac)

We found that folks who are across that spectrum love what we're doing. When people are in one, maybe there's like sub-strata of one, but one asks, many times, do you even have a CAD file of your assembly? I would say, depending on what industry you're in, aerospace and defense and automotive folks basically always have CAD. We work with folks that are, vertical and massive, right? So we work with folks in aerospace and defense, automotive, ag and construction machinery. Maritime I'd say the ones who are least likely to have a CAD file. They might just have a 2d drawing are Midwestern like ag and construction machinery like OEMs who like have You know, might be like a 50-person facility There's a 50-50 shot if they have CAD files if they have CAD then they are like fully You know fully cat it up. Everything is great If they don't they're like still working off of like a 2d drawing for their assembly It's very very binary And then in maritime, it's like very, very paper-based, which is unfortunate. We're trying to help push that a little bit more forward. ⁓ would say for the most part, everybody, there seems to be like a unanimous willingness to adopt what we have, especially if they are thinking like everybody's experiencing the workforce problem. Everybody's experiencing loss of tribal knowledge. So, you know, it's very obvious that what we're building is needed. ⁓ And so we end up being able to work with know, tier one through suppliers, OEMs, mid markets, enterprises, it's sort of like, if they have parts and they fit together, that is a customer for us. So, you know, we get to see a really broad spectrum.

Michael Finocchiaro

See you, David.

David Priev

So when it comes to like part manufacturing, it's pretty interesting because like the technology that you have today, cam softwares, ⁓ CNC machines, like all the technology around the cutting tools, there's a lot of functionality there that allows you to basically build your own rule-based automation, right? Like new CNC machines has so many... cool features in it, like tons of automation, tons of smart stuff, sensors, everything. CAM has like tons of capabilities that usually you just scrape like 5 % of its capabilities. ⁓ And the interesting part is that I would say if users and machinists and manufacturing folks were like highly educated about those new capabilities in the market, I would say a lot of them will be probably hitting like somewhere in around three. But unfortunately, because no one teaches you today how to camp program or no one teaches you today how to leverage all your CNC controller capabilities, so you basically end up by getting to the plateau of the organizational knowledge. And if you have some experts, that's great. They're going to teach you some stuff. But if no, like you're stuck. And then you see most of them are, I believe, around one to two.

Michael Finocchiaro

⁓ So then that does give you a huge opportunity, right? Because I was thinking and my thesis was that a lot of companies that are relatively low to the left of the scale, when they get something awesome like Dirac or Limitless CMC and they see what you can do when you've got better control of your data and less silos and CAD files, just even something as simple as that is there. When they, when you, have you ever had that feeling like when you put in your software that there's a sort of an aha moment, like, ⁓ that's how that's what the digital revolution is. That's, or is it a ripple effect? Or they're like, well, that's nice, but you know, I need a couple of years to kind of really understand it.

Fil Aronshtein (Dirac)

Yeah. 100%. So something cool that we've seen is oftentimes there's two big things that we've seen, right? When it comes to actually like, what does a company that makes a thing care about? They care about manufacturing, right? They care about actually getting a thing off the line. They care about getting a part out the door or an assembly out the door. And so a work instruction is very tangible to them. And so when they see that, hey, we have this 3D animated automated work instruction platform. And we say, yeah, the only input to this is a CAD file. They're like, ⁓ that's why we need CAD. yeah. And so in the cases that I kind of described where there's no paper waste, what's happened like several times is we have come in the door, they've said, we need this management says we need this. And that has actually catalyzed them to CAD up and like adopt CAD and CAD up their assemblies. And so it might take, you know, an extra couple of months before they could start using us, but we are like the catalyst for them to start adopting CAD. ⁓

Michael Finocchiaro

Yeah

Fil Aronshtein (Dirac)

Similarly on the you know, that's on the left side of the workflow on the right side It's you know operator consumption of instructions a lot of these guys, you know We do have like a PowerPoint or Word doc or Pete, know PDF export that you can print out but like we also have an operator view It's kind of the whole thing, right? We allow you an operator to interact with this 3d animated model and to do that you need a screen in some place in the shop floor you need a monitor or a tablet or something and I would say maybe like 30 % of the folks that we work with initially have some sort of screen on the shop floor. ⁓ And so they will often start out printing our instructions out. And then they're like, ⁓ we've been trying to get tablets in the shop floor. We've been trying to justify getting a monitor at these stations. Now we can. Now there's totally an obvious use case. So now let's go get a bunch of monitors. Let's set up these stations. so we've seen a really, it's like within two, three months of adoption, software. They go from being like paper to, you know, having monitors and operators with the operator viewer ⁓ because they're like, ⁓ cool. Now we can actually justify the investment because it's so obvious. Like we reduce error rates. It's like a whole thing. So I would say to the left, we are a really cool catalyst for some, you know, a little bit more archaic facilities to both adopt CAD and then also to adopt monitors on their shop floors. Because it's just like we're just, we are the

Michael Finocchiaro

Wow.

Fil Aronshtein (Dirac)

connection point that binds the engineering and the manufacturing workflows.

Michael Finocchiaro

How about you, David?

David Priev

So we had two crazy aha moments that we really remember. The first was when one of our users at one of the companies that works with us ran our recommendations step by step. And each recommendation, he's like, hmm, yeah, that's interesting. Maybe I'll do it differently. And then he accepted. And so we got another step and another step. And eventually, finished the program. He had like 30 steps in his cam environment in like 15 minutes. And at that moment he was looking, wait, I'm done. Like I actually programmed the part ⁓ and they were super excited about it. And that's what gave them like the motivation to try it again and to continue using it. And the second thing was actually we had a very interesting use case of a team that onboarded a CNC operator, the one that runs the machine ⁓ into the programming team. So he has no experience with CAM.

Michael Finocchiaro

Yeah.

David Priev

and we introduced him to our product or we call it CamAgent. ⁓ basically the CamAgent started showing him recommendations. He accepted them because he has nothing else to do. ⁓ And then he opens the recommendation, like he opens the operation that we just created in the Cam and then he goes through the parameters that we selected. And by that he starts to learn how to program. It actually helped him to figure his way out.

Michael Finocchiaro

Yeah.

David Priev

in ways way into the the cam environment and to the cam programming because you know it looks very intimidating when you like see cam or cat like it's like a it's like a airplane cockpit like too many buttons ⁓ but when someone like give you some steps at a time it really helps

Michael Finocchiaro

⁓ Well, we're almost out of time. ⁓ I really ⁓ thought that was ⁓ really, really insightful. Thanks, guys. ⁓ Any closing comments? What are you guys going to do? Are you guys coming into any conferences? Is any places we can see you in the next couple of months?

Fil Aronshtein (Dirac)

We've got some folks on our team over at Workboat right now. Workboat's a really cool maritime conference. If anybody is listening right now who is at Workboat, go see our guys.

Michael Finocchiaro

Awesome.

David Priev

That's cool. We're to attend four conferences probably like in 2026. It's going to be Hannover Messe in Germany and then two Siemens events, the Realize events in Amsterdam and in the US as well. ⁓ And lastly, IMTS.

Michael Finocchiaro

Very cool. I want to get to prove it. want to do that one. That one looks like so cool. Anything else? mean, that was really awesome. I think I learned quite a lot about CAM and about assembly. Actually, I wanted to say that I realized also that I was confusing part in assembly stuff. To your point, Filip, when I was writing this book on the history of CAD, I realized that I need to separate the history of how we did parts, which is really the PDM problem from how we did assemblies, which was more the PLM problem, right? It's sort of, you know, product structure and then all that other stuff was sort of that next generation that created PLM because then you needed change management, configuration management, because you were dealing with multiple parts. And then of course, on the machining is like, how do you go from CAD to a real part, right? Which is really cool. That's awesome.

Fil Aronshtein (Dirac)

Yeah. ⁓ Thank you.

Michael Finocchiaro

Uh, think we're, good for today. I want to say thank you very much to both Phil and David for taking time out of your busy day to, to tell us, um, give us your opinions and your, your perspectives on AI. Um, and I want to say thank you to the folks that joined in. And of course this will be on the YouTube channel and feel free to reach out on LinkedIn to Phil and David. If you've got more questions, think that David, you've got a couple of questions from that same, uh, Matthew Ferraro and the Q you might want to check those out. has some good questions.

Fil Aronshtein (Dirac)

Thank

Michael Finocchiaro

⁓ So thank you very much and we'll see you on the next episode tomorrow. ⁓ December 5th, I've got null space, so Masha and Misha ⁓ who from ⁓ Infinite Forms. So it's going to be more on the simulation CAD side tomorrow instead of manufacturing today. Be really fun. So looking forward to seeing you guys tomorrow. Thank you.

David Priev

Thank you, Michael.

Fil Aronshtein (Dirac)

⁓ Thanks for having us. Catch you soon.

Michael Finocchiaro

So how'd you like that? Was that good?