From Hypersonics to AI Workflows: How Ursa Major Is Scaling Rocket Production

0:05 [music] >> All right, welcome back everybody to the Additive Snack podcast. [music] I'm your host Fabian Alefeld and today I'm super pumped, super excited to welcome back Thomas Pomorski from Ursa Major. Since Thomas was last on the show, it's almost been a year now. Of course, a lot has happened not only in additive manufacturing has happened at Ursa. Uh a lot of news and interesting innovation coming out from that team, especially around Thomas from liquid rocket engines to hypersonics to solid rocket motors.

0:40 Um new facility groundbreakings and more. So, we'll talk about what's new at Ursa. But at the same time, of course, uh in the last year a lot has happened in the space of manufacturing, of additive, and in the space of AI. So, we'll also talk about how Ursa is applying artificial intelligence in their programs, how they see AI, but also other innovative technologies implementing and affecting manufacturing uh from their perspective. Uh and overall how that is going to change the way we build innovative applications.

1:16 So, therefore, as I said, super excited to have you back, Thomas. Uh welcome to Additive Snack. Hey Fabian, thanks for the introduction. I'm really excited to be here. Uh I think you're doing a a great thing for the industry and I'm happy to be a part of it. Yeah, uh so do you. Uh you continuously uh innovate, but you also talk about it and uh therefore I really view you as one of those leaders in the industry that helps us to to grow uh the space, uh to share innovative insights, and therefore really advance the industry forward. And before we dive into some of the really cool stuff that you're doing at Ursa, let's maybe take a few steps back. Like I mentioned, it's been a year, uh especially in our industry, a year is a lot. So, yeah, what has happened at Ursa in the last 12 months?

2:06 Yeah, absolutely. So, we've had a lot of exciting things happen over the last year. So, you know, Ursa right now is focused on three products. One is Hypersonics. The second is solid rocket motors, and the third is in-space propulsion. We think these three areas are really key key unlockers for the defense industrial base in the US. And we've had really exciting things happen on all fronts.

2:37 So, for Hypersonics, we have successfully flown over nine hypersonic missions with reusable 3D printed Hadley engine. So, it sounds, you know, pretty crazy to say, but yeah, we've flown our Hadley engine, which is an oxygen-rich stage combustion engine over nine times with our partner Stratolaunch. Um and uh some of those missions have been with an engine that's previously flown hypersonically. So, we've flown a mission, gotten the engine back, done this, you know, inspections, slight overhaul, and then flown it again. And these engines are, you know, roughly 80% 3D printed by mass on actually EOS hardware.

3:28 Um so, that's that's one thing. That's our uh cryogenic liquid oxygen uh kerosene engine. We have a storable version that we call Draper, and Draper has actually had two successful test flights with a program with AFRL in the past few months. So, we built a full vehicle ourselves and we integrated the Draper engine onto that vehicle. Again, that engine's primarily uh 3D printed. And uh the both of those test test flights were successful for were successful.

4:05 Uh on the solid rocket motor side, we've also been working on our links manufacturing process. We've talked about it a few times, but uh we actually have a full mix, cast, and cure uh facility uh at our headquarters in Berthoud, Colorado, where we're doing the propellant formulations, we're mixing it, we're doing all the manufacturing for the case, we're pouring the propellant, and we're test firing them all. And we heavily use additive both on the tooling side, but as well in in the actual cases to uh enable flexibility. So, one of the biggest problems right now in the US is that we're depleting a lot of the munitions. So, um we need to have a new manufacturing process around solids that enables scaling different type of products, but don't require that fixed tooling. Right?

4:57 And that's what additive is so great. So, so we're building this unit cell approach around our solids that we can build any size, any type of motor, and then when we qualify the unit cell, then we can scale that to multiple different sizes, different platforms, things like that. So, uh we we've had multiple test flights successful of of solid rocket motors as well. Uh so, yeah, it's been an exciting year on the product side. And then on the additive side, we've had an equally exciting year. So, we've uh gotten multiple EOS M 450 machines, which we had announced uh a few months ago uh with you guys.

5:33 Uh we're we're running those machines. They are amazing. They have four end lights in them uh with upgraded gas flow and and optics and things like that. And they can really uh crank out with material. And then on the uh the the uh slicer side, we've been really uh working with with Dendrite and with folks from EOS to develop what we think is one of the most advanced uh slicing pipelines in the entire industry. And we've uh been printing hardware with that for the last few months, and our hardware is being hot fire tested and uh actually being put in applications as we speak. So, we're continuing to refine that and build it out. Uh we've learned a lot, but yeah, it's been a lot of uh a lot of exciting stuff over the last year. It's awesome. Yeah, you guys are definitely not uh not slowing down. Uh you guys are going you're going hypersonic yourself.

6:28 Um That's right. Uh uh let's let's maybe before we go deeper into additive because I want to ask uh ask you a lot about those uh those especially those software innovations that you guys are working on. Um maybe double click on, you know, hypersonics uh versus uh versus solid rocket motors. Um What are some of the uh the the key challenges of hypersonic propulsion, right? It's one of the hottest things in the defense sector at the moment. What are some of the challenges uh of of of that application and how is additive helping to to solve those?

7:03 Yeah, it's uh absolutely. So, the biggest challenge to hypersonics is building affordable production system. So, like it's not that hard if you have hundreds of millions of dollars to build a hypersonic system. Like you know, we've been doing it for for years um as a country, but those are may basically test vehicles or very bespoke applications or extremely expensive. So, um the real real holy grail is can you build an affordable manufacturable hypersonic vehicle? And there's a couple different um there's a couple different avenues that you can go down. So, there's some air-breathing systems that are using uh like uh you know, uh uh high-end production like a ramjet or a scramjet. Um Some other exotic propulsion systems, uh rotation detonation engines, things like that. Uh they're air-breathing. Uh there's some potential solid applications. And uh what Ursa thinks is a unique application is we have a liquid uh a rocket a liquid rocket engine version, which is around our storable draper engine. And so, what's nice about this engine is that it enables a lot more flexibility in the uh profile of the flight in that um doesn't need oxygen. So, cuz it carries uh the oxidizer, which in this case is hydrogen peroxide, with it. Um that also means that you don't need a really sophisticated ignition system because a lot of as as I'm sure a lot of people know, a lot of the more exotic propulsion systems have to already be traveling really, really fast to get started like a like a scramjet or a ramjet. So, you have to now build a vehicle, you have to be able to test the ignition in supersonic or even potentially hypersonic regimes using one of those systems. So, now you can imagine how expensive that testing is, how difficult building all those systems are. So, what we're trying to do is, "Hey, we have this liquid rocket engine.

9:27 We can test it in any regime." Uh we've already flown hypersonically. We know how to how to use an engine there. You can test it on the ground. Um and we know how to build additive hardware uh affordably. So, we really think we have a novel approach to solve what I think is the biggest challenge in that providing a low cost, you know, in the low low million-dollar range uh full vehicle that can enable uh hypersonic applications.

10:00 Interesting. Interesting. >> [snorts] >> And you're you're therefore also increasing the ease of deployment, right? Because you don't have to start in in flight and therefore do an ascent there. >> Exactly. Exactly. Yeah, so so you know, we can have a booster attached to increase the range. Um but there's no really requirements on starting it. So, um you don't have to you know, you don't have to air launch it though you can.

10:30 Um so you have a like you said, you have a longer flexibility. And then it always it'll lower the cost to qualification and production as well because again, we don't have to validate the operation in these super challenging regimes. Like I mean, to to get to test in a wind tunnel that's supersonic is very expensive, right? So, you have to make back all of your R&D cost on the So, um you know, there's a lot of competing ideas, a lot of competing systems, but you know, for any of them the end use vehicle full vehicle has to be low cost. And and you know, what that exactly means, like I said, you know, probably on the order of like a million dollars a unit, but they can't be 20, 30, 50 million dollars and right now a lot of the bespoke applications that we have in the US are you know, test vehicles, target vehicles and those are talking tens of millions of dollars per per flight.

11:32 Interesting. So, [snorts] um when it then comes to to your engines really, it's interesting because, you know, for the longest time, if you think of additive, people always think of an expensive manufacturing technology. Right? And now we're actually at the point where additive allows you to be more cost competitive than probably conventionally manufactured engines, which is also to me again a huge testament to the to the technology. And now if you talk about solid uh rockets uh motors, I mean, here you have an even bigger cost pressure, of course, right? Cuz now you're competing with uh with uh with with cast material. What does that whole production uh uh chain look like, and what exact role does additive play? You mentioned a little bit about uh using additive in in tooling.

12:20 Um where does additive add value to the solid water rocket motor production? Yeah, so solid rocket motors have an extremely competitive uh cost profile. Um but the challenge around solids right now is that they're all built on fixed manufacturing lines. So, a manufacturing line can make let's say 413-in motors a year. Well, if you want 1,400 or if you want 401 motors, you have to build a whole new production line.

12:55 So, that production line's going to cost you a few hundred million dollars. So, you know, yeah, they'll be cheap when you get to that rate, but you have to invest that capital, you have to to amortize it over time, and you have to make that commitment up front, whereas maybe in, you know, 2 or 3 years after the production line's built, maybe we want a different product, or maybe the industry has innovated and that product's no longer relevant. You know, a good example as to this is the use of small cheap drones.

13:26 So, so maybe things pivot where we don't want these like larger, more expensive solid rocket motors. Maybe we want a lot more smaller ones. Well, now you just built the production line which took you 3 or 4 years to now produce, you know, the larger ones. So, where Ursa's really focusing on with additive, right? Because one machine can build, let's say we're building a a section of a solid rocket motor. Um we can build a 4-in, we can build a 10-in, we can build a 14-in all on the same machine.

13:58 So, if you go to what we're calling a unit cell approach, where you develop all of your tooling such that it can handle multiple different sizes, then it you can set up your factory for um one size and then you can basically just change the the digital manufacturing requirements and the the product design and then you could uh build a different size motor. So, that's where we think a lot of innovation can happen in the solid space and that's what we're we're focused on addressing. And on the additive side, we do it both for tooling as well as actual end cases um as certain materials. So, um you know, there's certain features in solid rocket motors that are really complex that are typically they're either um uh forged material or cast material and the lead times on those are extremely long.

15:00 So, again, if you're setting up a production line, you do a large order, you know, the material will be here in like a year. Like the the lead times for the more exotic materials are or about a year. Um there's some like high-strength steels and things like that. Maraging steel, Armat 100 is another big one. Um you know, so the other thing that we can do is we can go in and get these materials and we can build these things super rapidly and then also apply versus philosophy of design, print, test, iterate. So, we're kind of we're kind of addressing multiple different challenges all at the same time with our design and our manufacturing process. Interesting.

15:37 Yeah, so you really have two sides of the coin on your Draper and Hadley. Of course, additive is also really leveraging you're leveraging the design capabilities of of the tool. Whereas in in your Lynx production, it seems really mainly a supply chain case. Um I'm guessing that there's not any unique geometries in your engines that leverage additive or are you still including some of those potentials uh there that >> Yeah, oh there's still there's still are challenging geometries. Now, I mean we're we're we want these to all be drop-in replacements for for what's out there. So, we can't do anything crazy.

16:12 We have designed and hot fire tested crazy designs that utilize additive, you know, double wall things with insulation and like lattice structures and things like that. And we've you know, we've tested them and they work. The industry is not ready for that yet. Right now, the industry wants is drop-in replacements for the the platforms that they already have and then I think in you know, 5 years down the road, then we'll be ready for more um advanced solutions. But, you know, our again our process is agnostic. So, yeah, the parts we're printing are I mean, they're still fairly challenging.

16:44 Um but, a lot of them could be made conventionally. However, again, you have to build in factory. You have to build a production line. You have to pick what you're going to do. And that does not allow flexibility and and as the the world keeps evolving, I mean, just think about how much is we've done in the past year. I mean, all of that of manufacturing is evolving so fast. So, um to me, I think there's a lot of value in building that flexibility in from the ground up. Again, cost if you factor in the entire cost is fairly competitive.

17:17 Uh you know, the unit cost is going to be a little bit higher, but you don't have to to have a full production line factory cost cuz you can also build each unit cell, you know, organically as you need uh, the production, right? So, so when you're investing CapEx in in a facility, you have to build the whole facility before you can get a single motor off the line, right? So, in here, we can build one unit. So, we can get low rate production. We can do qualification testing. And then we can just boom, boom, boom, duplicate the unit cells uh, as we need to scale. So, also helps spread out the capital as well, which is good for startups and VC funding. Yeah.

17:54 Now, it it's very exciting, uh, moment in time, right? Where we where we see additive contributing to those types of programs that I think, uh, even 3 to 4 years ago, we would have not thought that they would be, uh, even close to being cost competitive, but now the technology is mature enough, it's reproducible enough, um, it's scalable enough where uh, you see applications like yours, you even see, right, uh, consumer good applications, uh, where additive is cost competitive to conventional technology.

18:24 So, it's a very exciting time, um, just from a just from a hardware readiness perspective. And then, as we discussed earlier, right, now, uh, we're starting to see slowly and slowly AI creeping into into our fields, right? I think if we compare it to other industries, we're still at the very, very early beginnings in in additive and uh, you know, if you watch, for example, Duann Scott, he tests every tool that's out there on a, you know, prompt to design basis. Most of them fail, uh, but some are actually impressively good, right? Um, cuz he always uses the same, uh, geometry of a monitor mounts, uh, that, you know, should be somewhat simple, but, uh, most of most tools fail.

19:10 Uh, then Autodesk, uh, announced, uh, a prompt to parameter optimization tools. So, there's a lot happening. Now, even Cloud, uh, came, uh, came into this space, uh, with a collaboration with Fusion. There's probably even more that I haven't even, uh, even seen. So, uh, let's maybe, uh, start there. So, where do you today see value of artificial intelligence in the manufacturing {slash} additive manufacturing sector?

19:41 Yeah, so I use, uh, AI every day in my additive development. Um, and so right now it's already useful. Uh, so where I use it is I mostly use Cloud Code CLI tools. Okay. And I'm working on two main things. So, one is, you know, we talked about, uh, laser development. A, you can use it to develop algorithms and design of experiments and, uh, new functions very rapidly that we can go and test out. So, I might say, "Hey, I read a paper about this new scan strategy.

20:19 Maybe it's, uh, a spiral scan or maybe it's island scanning or whatever else." I can go plug that paper into Cloud Code. I can give it our slicer repo and say, "Hey, go and build this function in the in our repo and then I'll go test it out. So, they'll go go and build it, you know, uh, I'll verify that it works. I'll slice it, the new algorithm, and then I can go and and test the algorithm on our printer. And I can do this rapidly and iterate. And then once we find something that works really well, then, um, you know, we have a software team as well and they can go and um, they typically rebuild things because, you know, the tools right now, there's as anyone that uses them knows, they uh still aren't mature enough to go into a critical environment that we have and you can trust everything that they do.

21:18 So, um you know, we've been rebuilding all those things, but that fast iteration, which is really what we're after. You know, I'm really focused on acceleration. How do we do more? How do we go faster? And this has really enabled us to screen for new ideas, new parameter strategies, new scan strategies um in a really really quick way so that we know things we need to focus on. And then on the back end, I think so that's on the front end of that workflow. On the back end, the other thing we've using it for is to help a lot on the data processing, data management side on the additive. So, uh uh we can go through this later, but I built a full um data fusion workflow all with AI where I can pull all the data off of our EOS machines, put them in a database, I can view the data.

22:12 Um I also played around with defect detection. Um and then uh powder bed images, uh live camera feeds, all that stuff I built from the ground up with AI. And then what I like about that is again, I can test all these function all this functionality out, see what we want, see what's useful. I can show engineers it, say, "Hey, what's useful? What do you like about this?" And then we can go and spec out the software uh that we actually want to build and and makes it a lot easier to deploy this to the software team where we already have a full framework built out. Here's what we need. Here are the requirements.

22:49 And then they go and and polish it off in a very robust, you know, production-ready way. So, those are two cool applications. Wow. So, you really I mean, you're using AI to prototype in different aspects, right? AI to prototype really new parameter strategies. Um but also to prototype software that then your software team actually writes from the ground up. Probably still with AI help from Cloud Code or Codex or whatever, but but in a more mature non-Vibe Cody way I can I can imagine.

23:28 Exactly. Yeah, I mean when what I'm doing it, you know, I have I have a day job which isn't isn't coding all the time. Um so I'm I'm kind of doing this on the side. Um so, you know, I'm doing large projects with AI which then adds in all those type of vibe code vibe coded defects that we love to laugh about on the internet. Um so it's great for development for the way I'm using it, but then, you know, the software engineering team, they can use it for very very small things and get it very directed.

24:00 Um instructions and that works really really well to more rapidly develop more mature software, but um I'm and and and we'll see I have a a good example that we can go through and hopefully it'll work out, but you know, I'm giving it very broad instructions. I'll give it a lot of details about what I'm after, but then I let it go off and do its thing. Um and again, great for prototyping development and and just increases that that development speed which is what I'm really after. Yeah.

24:32 When you cuz it's not been that long ago. You remember pre pre-Cloud and now post-Cloud. Would you how much more efficient would you say these tools make you? Like would you say A that a lot of the development work you're you're doing you wouldn't even be doing because it would be Jeffords? And would you say you have now twice the outputs, twice the product What does What is that X uh look like on your end? Yeah, I mean it's hard to even put a number on it. I would I would easily say 2x or 3x the amount of like DOE and experiment output um because of this capability. Cuz I don't like this is all stuff I can do myself um but I'm super busy. You know, we're building a team, we're building machines, we're qualifying hardware, we're we're getting feedback from hypersonic flight, we're doing design for additive work with engineers, we're developing new workflows. Like like that's my day job and then you know, on the side I have Claude running and you know, and Claude agents running and then I'm like, all right, let's build this. Let's do this DOE. Like let's Here's this paper. Like what do you think about this? And so like I can use let's say 5% of my time and do the work that a a full engineer could do um on the side of of my jobs. Yep. Um yeah, it's it's it's really amazing what it enables and this is just the infancy. You know, we're still using a fairly outdated version of Claude because it um has to be compatible with AWS GovCloud uh which is like three or four versions behind. So we're using like you know, uh 4.5 and and now they're you know, all the way up to Mythos preview for some people. So um I think there's a lot more potential and as these things continue to evolve >> Like we're going to be able to use this more and more and more. So what what I say now is like people need to start figuring out how this fits in the workflow and either if it's not directly useful on the production side like I really confident it's useful on the development side and and people need to understand where where it's at and and like the level of maturity it's at.

26:50 Um like right now cuz like I said, I'm using it every day. It's already useful and I'm Hopefully we can do a demo and and we'll see it see it work in real time, but It's only going to get better. This is still like the infancy of the capability, I think. >> it's ever going to be. Uh, yeah. >> Yeah, exactly. >> If you can if you could show, uh, that'd be that'd be awesome. Let's let's uh let's look into some of your workflows. And while you do that, so how many people on your team, uh, have the same, uh, approach?

27:22 Right, so I guess my question is if you're a five-person team, everybody does what you're doing, you're now acting like a 15 to 20-person team, right? You have the same output as a 20-person team. >> Exactly. Yeah, we have a very small team. Uh, there's just uh four of us on the development side. Okay? Um, and we're we're outputting, I mean, like we're a team of at least 10 or more with the type of workflows that we're able to Yeah, to develop. And it's actually even more because of, you know, the experience level that myself and some of the folks on my team have, right? Like it it the AI can almost be a multiplying factor if you already have the knowledge. Cuz if you tell it exactly what you want, it's really really good at going and doing those things. So, like if you're an entry-level engineer and you haven't had that much experience in there, you're not going to get as much benefit out of it because you don't know all of the nuance around the type of factors that could be involved in the thermal profile, the thermal history of a part, like all those things. But if you can tell Claude, "Hey, these are the things I'm interested in. This is why I'm doing it. Like what are all the things that we should change?" And then they'll come back with all these ideas and say, "Oh, like this is the thing you should go off and do." And so it's like as you get more experience in the additive and then you can go and apply these technologies. It's like it's almost like a an exponential function of productivity where like you could have an army of you know, less experienced engineers and one person or a few people with these tools can do, you know, more work than an army of people. Yeah.

29:06 You know, you're 3x in 6 months or 6x, right? Um in 12 months >> Exactly. Claude is going to Claude is screening papers for you, identifies which ones are worth prototyping and will say, "Yo, Tom, I got three new three new strategies. Which one you want to print?" Exactly. Very interesting. >> So, it's important to build those workflows now. Yeah. Yeah. So, this is going to be a really cool actual snapshot into how our slicer operates. So, this is VS Code environment and we're using Claude Code CLI integrated with that.

29:49 Um I've already pre-written uh some of the prompts, but we are going to run this in real time so you can see one of the actual workflows that I might do. Um so, I'm um prompting I opened up a new Claude session and I'm prompting it um telling it it's a an expert software developer for laser powder bed fusion machines working on our slicer repo. Uh read through what we call the Polaris slicer and the build setup template and explain to me how it works. So, this is good for for a brand new instance of Claude to come in and uh understand the coding platform that it's working on and get everything it needs in the context. And you can do this and create what's called markdown files to actually um give Claude more or less notes on what it's doing and what it's read. So, it's going to go through and read through the code and it's going to tell me, "Hey, here's how the this the the slicer works. It it loads in geometry, uh has different uh parameters such as layer thickness, parametric angles, down skins, up skins, hatch parameters, uh contours, um and then uh the configuration is driven through what we call TOML file, which has all of the input parameters.

31:11 Um and then uh engineers can change laser parameters using these setup files uh without actually going in and editing the code. Um so, what we're going to do now is I'm going to tell it the folder that I'm working in. So, we're working in uh the EOS AI podcast demo. I I'm telling Claude, "Hey, go take a look at this. Look at the set the generic setup file I have in there. Understand how it works with Polaris architecture.

31:38 And then think about how we can use these functions in Polaris to make a part name specific segmentation or hatch group strategy. So, the way that our code is worked up is we can name the parts certain things like sample one, sample two, sample three, or and you know, any name that we want. And then we can give each part completely different parameter strategies automatically using these segmentation configs.

32:09 So, now I'm telling it, "Hey, here's the parts already put in this folder. Here's the generic segmentation." We just have defaults in there for now. Um and it's going to it's going to come in and learn all about that. You're setting up Do you have any questions? >> So, you're automatically setting up a DOE in this case with unique parameter strategies for each uh test specimen uh without having to do any manual, um, parameter uh, application to the individual parts, but Cloud does it all for you.

32:45 That's That's exactly what's going to happen. Yep. >> Okay. So, I So, sometimes it gets it. Here's one of the things that we had to try. Sometimes it knows the parts that are already in there. It didn't get it this time, so I'm going to tell it, "Hey, there's already two parts in this directory that we can use. What are their names?" And then I remind it that the directory is there. And then it's going to hopefully come back and say, "Here, here's the two parts in there.

33:10 Uh, now we can go in and we can change the parameters. Yep, so here it is. So, we have R D A M rec 126 number one and number two. And so, we can go in and have a build strategy for that. We can have different contours for that. Um, based on what we want to add. All right. So, um, what we're going to do in this demo is we're going to try a different hatch pattern strategy. So, a lot of people might use chess or might use stripes. We call it checkers. We're We're basically tiling, uh, core, um, uh, in the part. And we're doing this against gas flow. So, we have an algorithm set up where we define the machine coordinates, and then we scan into the gas flow, and that's how our standard checkers works. So, it's pulling that out from the file right now. I say, "Hey, here's type checkers.

34:11 We have 10 mm by 10 mm with uh, 0.12 mm overlap." And it's going to show you the difference from an M400 platform, how it works, and why it matters through all management, you know, scanning into the um, and usually it gets excited and wants to do something, so I tell it don't do anything yet. Um but now what I'm going to do is ask it to create a new scan strategy. So, I'm going to say uh let's keep the default checkers on part number one, but I'd like you to apply a new strategy to the second part.

34:51 So, the new strategy should sort tiles from the center of a part uh from the outside in. So, if you have a certain certain profile or you have uh certain types of features that you might want to do on the part um we can now, instead of sorting to the machine coordinate frame, we can sort to each individual part that's on the frame. So, it's going to come in here and it always likes to jump around here.

35:18 Um so, I tell it, "Hey, um here's what I want you to do. Here's where to go look in the code. Um I would like you to to plan this task out. Here's what I would do. Create a new hatch pattern function in the slicer. Sort the tiles from the outside of the part. Uh edit the segmentation file with this new feature and apply the hatch pattern to part two using the tunnel file." So, it's going to come in here and um sometimes it gives me options, sometimes it tells me everything I need to do. So, this looks hopefully good.

35:54 Um in here and import that. I'm today four solution and I will test it. Uh And so, now it's going to go off and it's going to edit the code. It's going to edit the setup file with the part names that are already in the folder. Um and we're going to give this a few minutes um uh to to do and then we'll we can check back on here in a minute or two.

36:31 Um when hopefully it is done. So then um So how do you work then as a team? If If one of you guys uh develops a workflow like that, uh do you have a shared prompt library uh that you guys share within the team for different uh different workflows um or or how is that then disseminated across uh the team so that you guys don't reinvent the wheel uh and you also, you know, improve and build upon each other's uh AI work here.

37:01 Yeah, so we have multiple GitHub repos that we use. So we have obviously GitHub repo for our our slicer. We have uh GitHub repos for different setup scripts and things like that. And then we have um internal wiki that we run. So it's, you know, similar to Wikipedia, but uh we have it internally for um you know, less official like more like helpful things. So um we put a lot of helpful things around novel workflows that we create and uh things that can help.

37:42 We like to think about it as like, "Hey, you have a new engineer you have someone that wants to learn this. How do they do that?" So we try to set it up as like an educational resource. Um but I mean, we all sit next to each other, you know, it's one of the benefits that we all try to to be in the office cuz we're also like building stuff and we like to see our hardware and things like that. So uh we also just talk about it a lot and and play around with things. But um I really like the the the wiki resource i- is pretty easy to to host yourself and um yeah, with with the AI tools being able to help you as well edit that. Um but it's easy enough that most people could do it on their own.

38:19 Um that's a really good resource cuz then it's easy to update too as things change, cuz even like I would say, you know, 6 months ago, a year ago, like it wouldn't be able to do what we're going to do here today. Yeah. A lot of us a lot of us is changing. Um, I think it's still it's still uh still thinking. So, let me let me ask you this then. How far away are we from uh a production deployment, right? Cuz now we're talking uh more like we said, DOEs, prototyping, development work.

38:55 Um, when do you start to see value of AI in production environments? And that could be a like a true AI develop process parameter that you that you qualify. That means qualification work itself of a new of a new program and accelerated work. That might mean um uh process develop what develop parameters based on uh required material part properties, right? Like how far we are how far away are we from from a future like that? At least where it's maybe in a guidance environment, just like you're you're getting guided when you're you're coding uh today.

39:37 Uh I think we're still a fairly long way from that. I think there's I think the challenge with doing AI on like Let's say you're defining your thermal All right, I have a good example. So, there's there's a recent paper that came out. I don't remember the name off the top of my head, but basically they used a simple thermal FEA models to train a neural net uh model >> surrogate model to basically five simulate a part.

40:16 >> Mhm. Mhm. Right? So, um and it reduced computation by like, you know, five orders of magnitude or something. Um the problem with that is now, how do you validate that there's not some edge case. How do you like cuz you don't actually know what it's doing, right? I mean, you have some idea of what it's doing and you know the patterns and things like that, but you don't have a good insight, especially into the edge cases.

40:45 So, I think that we need functions to better qualify how the code is functioning, so that we can perform a lot more like unit testing on certain aspects of the code versus just testing the output of the code. And I don't and and and I think that's going to take a long time before we actually have like ways to go in and and validate those sorts of things. This is So, I don't know. If you had to put a number on it, I would say 2 years, but like it's going to be able to do those things probably in the next 6 months.

41:30 Um but it's going to be really, really hard to validate. The validation's way longer than it is to build all of these these tools. Then that that really also still means a huge benefit, right? Because even if it's saying, "Oh, yeah." Even if it's just a FEA surrogate model, um the but it can tell me instantly while I'm designing a new part that I'll have distortion here and uh and issues there, right? I have immediate feedback. I can design something very close to uh perfect. Now I run my final FEA to confirm uh to confirm this, uh and I still have saved a lot of time and I probably still have a better a better product that will print pretty much immediately without issues. Uh cuz I've uh I've I've I've gotten continuous feed feedback throughout my engineering process.

42:20 Yep. So, um let's jump back in here. Um So, it just finished. Uh what was built? So, uh the new feature sorts tiles based on each client's centroid rather than global reference point, enabling part-specific thermal management strategies. So, it came in. It uh added a new function that identifies a part uh coordinate system in our hatch tile sorting function. And then, it sorts to a different sort point. So, now there's two parts in that folder. One uses a standard checkers strategy, the other uses the checkers outside-in sorting. It automatically added the setup file, added the parts. I didn't have to do anything. I've only done exactly what I showed you guys.

43:05 And now, I ran the part. So, it imported the two parts. There's two 20° cones. >> Oh, cool. Um printed, they're labeled here. Um And one thing I didn't show you that's also built in our workflow, uh it automatically generated the support. >> Wow. Um if you color the surface emerald or uh pink purple color, it automatically builds supports up to that surface. So, all I did was click run on the setup sheet. It automatically imported these parts, applied this new strategy that uh Cloud Code came up with, and we told it to run, built the supports, and now it's slicing for our EOS M 400. And hopefully, the output looks good. It's going to take about 10 minutes.

43:50 Uh we'll double-check the output, and then we'll try to actually start it and run it on the machine so we can see the algorithm we developed just in the span of portion of this podcast. That's awesome. And that's probably, especially if you have a full plate, right? That's a days of engineer uh of engineer work. Oh, yeah. And I could I could have asked it to try anything. I mean, I could have Yeah. I could have developed an equation for thermal management around the 20° angle, and I could have put you know, normally I run a little more parts, but I just have a couple on here, so this is is simplified, but I could have a whole plate of cards. Every part's named differently. I could just have a giant Excel sheet that's not formatted or organized or anything. And just things are kind of labeled. I could tell Claude, "Hey, make a setup file with this sheet." It'll make that file. It'll give each part different parameters. And then I can just go and run it, and it takes me 10 minutes to do the most complicated DOE you can think of.

44:48 Easy. Do you Do you see a future where um you go through this process? Now you print it. While you print it, uh Claude has access to uh the OT thermal profile. Detects that some parameters do create overheating. Adjust them on the fly. Right? So that after layer 1,000, you've actually gone through as you go through Z, you've gone through even more iterations of your parameters, and you end up with a with one one winner.

45:21 Yeah, absolutely. I mean, and and And to build on like where this is going next, and it's absolutely the ability for Claude to now go out of the coding environment and start interacting with other types of data streams. So like, you know, one of the new tools that's um coming out or or is out already is the ability for Claude to control different functions on your computer, to control you know, different programs, and to to work uh in different in different contexts. So, um yeah, being able to go in like the ultimate would be, "Hey, go find this program, understand how it works, pull the data, figure out how to interact with the data, and then now compare it back to the slice that we just created.

46:05 So like we could do all that now, but it would require a ton of my time to go in and say, "Okay, here's where the data is. Go here, pull the data." Like I'd have to do it very manually. The real big unlock, like the next like step change, is when I can just tell it what I want. "Hey, let's go compare this to to OT. Here's the machine that has it on it. We already ran the build. Go figure it out how to do it. And then once we do that, I mean, it's going to be another you know, 10x unlock in productivity to to do that.

46:32 It's probably even possible today with Open Claw. That seems a bit ris- Yeah, I'm not I'm not installing that on my work computer, that's for sure. Uh Um so, while we're waiting for this, um I want to show you another thing uh that I've been working on. Cool. So, um this is uh the EOS web monitor that I built.

47:03 Um we'll reload it here. But, this was 100% built with AI. So, no uh no real software engineer uh has helped. Um but uh we're we're rebuilding this um for real now. Um but I was able to prototype everything that we wanted and all the different functions, things like that. So, you know, I got everything working and I and I went through a bunch of different iterations cuz it's like we don't really know what we want. Um so, what this does is it uses EOS's REST APIs.

47:40 It um pulls all of the data from the machines, puts it in a SQL database that it generates on the machine, and then projects that data to a web page that it built um that we can get to through this IP, and you can see all of our machines are named after bears Um cuz Ursa Major. Um just pretty funny. So, right now three of the machines are printing. One of them uh is ready for us to um start a build on after the build's done.

48:16 Um but you can see we also have a live video interface of the machines. >> Wow. So, um on the right that's our M290 flex. Uh we're doing uh tensile bars. We're doing a lot of heat treat studies um to look at lowering the cost of heat treat. So, um we're just running simple tensile bars and we're doing different shared runs to try to um lower the cost by increasing the flexibility of the heat treat, seeing what it does to grain structure, or tensile properties, things like that. Um so, as we have time in between parts, we fit those in. And on the left we're printing some components. Can't really talk about that.

48:53 Um and then the machine in the middle Blue is resting. >> Uh like I said. Yeah. But this is a live video feed of all of our machines right now. And this is all built >> Wow. like I said, with AI. We also have what's really cool is a uh email handler as well. So, um we have status reports on our machine. So, every um hour it sends us a status report, what layers the machine's on, build platform temperature, chamber pressure, everything, how many's printing, how many's idle. And what's really cool is when there's a build interrupt, or there's a stop, there's an issue, we get an alert. And so, we can set up the filtering in the email to to have a folder that gets our status emails, but then uh goes, you know, notif- notification when there's an interrupt, when there's a machine issue, and then sends you a snapshot uh in real time of what the camera's looking at in the machine. So, if we have a recoater snag, or uh short feed causing issue, or something, we see a a an image of what's what's actually going on in the machine.

49:59 Connect the Okay, wow. That's really cool. That's really cool. Yeah. How long did it take you to build that? Uh it took me like a week. And like I said, I'm only doing this as like, you know, 5% 10% of my job. So, like and one thing I wanted to bring up and part of the reason I want to show everyone just this show was possible, but the real reason this was was possible was because of EOS's really good API documentation. So, this is kind of my call to the industry and also, I mean, also to EOS as well.

50:36 The real unlock in the future is going to be the companies that have really good documentation for how do these tools interface with the real world? Or with carbon >> or with machines. So, a great example is the Connect Core API that EOS has developed. There is really good documentation around that. So, I all I did was I took all the documentation, I fed it into Cloud Code, I said, "Hey, here's all the documentation for this. What we want to do is build a, you know, database that can query the machines every minute, every 30 seconds, whatever, and then pull the data off. Here's the Here's the information I'm interested in, powder in the dispenser, um you know, what layer the job's on, when it's going to be done, all that stuff. And then, you know, build database for that. And then, once I did that, I said, "Okay, you have this database, now just go build a front end web page that anyone can view." And it can do all that, you know, on its own without any prompting. But, um so so one of the things that I really wanted to talk about is that uh one of the key unlocks for like the next phase of, I would say advanced manufacturing is any any of these systems that we can create really good documentation that the AI can then use to ingest the data or to interact with the data.

52:01 I am totally right. I think um system access, system openness uh to allow I would still call you a power user today, right? You are definitely on the bleeding edge of of additive. Um but really that means you're an early adopter uh and that means the early majority is is next, right? It's definitely key to scale scaling the technology and and uh you know for companies like yours to use additive to achieve a cost advantage um you're only able to do that by efficiency gains and by really pushing the limits of the technology. And this seems to be really the next frontier of pushing the limits of of tech. But before that was, you know, system reliability and and new lasers, but but that's all matured. It's going to continue to improve but not at a rate as as tools like AI are going to improve the the tech, I think.

52:59 Um and as you talk about advanced manufacturing you must forgive so what does that mean for a actually advanced manufacturing cell in the next two to five years, right? Right now we're we're mainly talking still connecting different tools within the additive ecosystem, but you could then also include post machining, you can include inspection, you can include other aspects of the of the process chain. How automated do you think um the full production chain is going to be potentially even including robotics um with the help of with the help of AI really.

53:41 So it's funny. I think I think robotics will kind of be one of the last things that that gets deployed, but I'm going to I'm going to I'm going to shove that one here till the end, but um I think this whole process is going to get extremely automated over the next year. I think that's going to be the name of the game. So, let me just hit on a couple things uh that we're working on right now and that are almost ready and can be rapidly deployed. So, good example is like uh build data reviews, right? Um there's some software out there that already um can do that. Um there's uh companies that have built algorithms, but with these tools right now, you can just build them yourselves. So, I I trained a uh an ML model with Claude to detect short feeds, for example.

54:28 Um and so now, eventually, when that becomes more robust, now we can just automatically review every build and flag any short feeds and you know, the operator can come in and check that. Um and then [clears throat] the the near the the further future state is it just checks everything it needs to and says, "Yep, it's good to go." And then you don't even have to look at it. So, um I think there's a bunch of different steps. So, the first step is how can we um do more with less people? So, a lot of that is is AI flagging flagging data that's anomalous, flagging defects that it finds during builds.

55:12 Um and uh that enables you add more machines without adding a lot of additional quality personnel or additive engineers, things like that. That's one key unlock for additive, I think, is that when you build to add machines, when you build to scale cost-effectively, which means you can't have engineers interacting with every single piece of this. It's got to be automated. So, It's got optics without the optics following the Exactly. Exactly. So, so we we have automated data processing, automatic flying things. I think, you know, the automatic acceptance is, you know, kind of we talked about. I think it's going to take a long time to to really have true validation of that.

55:54 But, like it's already ready now to like build these workflows and build these pipelines. And the other thing that I think is super important that not not a lot of people understand is that this data and the ability to view the data need to be integrated with your manufacturing enterprise system as well. So, like you need to be able to tie all of your components that are serialized and have qualifications and all that back to all of the build data and all the information and you have that feedback loop. So, another another key thing toward that automation is to have all of this data in a pipeline that's easily accessible.

56:40 So, I think that's going to be when we talk about automation, like it sounds weird to say that the operator is going to view the data, but the key is that it's easily accessible and it's easily all tied together so that you don't have to spend a lot of time doing a task of it's like, "Okay, I got to click through folders or it's on SharePoint or it's on a shared drive or it's somewhere like like it's just in a database and then you have a you know, front end that you can either build cloud or there's commercial tools as well, but then you can go in and all the data is there and it's interactable the way you want to interact with it. That's the huge unlocker and then eventually those things will go to automatic validation, you know, kind of like we talked about before.

57:23 Yeah, you're and definitely you're even not necessarily replacing humans. You're just again aiding humans and these humans manage different call them agents that's uh then the human still validates the final uh the final outputs and confirms it. But you're not taking the human out of Yeah. in that regard. Well, and it's it's it's kind of goes back to our our previous comment where you're 2x or 3x-ing the productivity of the engineer. So, now instead of the engineer wasting their time finding connections between data or what build was that in or what part's serial number is that?

58:02 Instead of the engineer wasting this non-productive time, now it takes them 2 seconds to type in the serial number and then it pulls up the build, it pulls up the data, images. It already flags where there's a potential problem. And now the engineer spends, you know, a minute going in, boom, oh, there was a short feed here. Oh, there's a recorded collision. Here you go, you know, put a discrepancy report on the part and then move on to the next task.

58:29 Yeah, and therefore you also I mean if I if I think about what you're what you're saying too um but there's there's a big debate on you know, is the manufacturing sector growing in the US, yes or no? Uh right, uh there's I think different there's different voices uh if you listen to economists or um not even politicians uh but uh there's not a lot of consensus on if manufacturing is is today growing in the US.

58:59 But if I listen to you and if I listen to really especially our industry there is no doubt that advanced manufacturing is uh is rapidly advancing in the US and that is not due to tariffs, it is really due to in this case artificial intelligence making processes more efficient allowing startups like Ursa to act like very large companies uh and producing similar output. But also if you think about it, um AI moved funding away from software companies or SaaS companies and applying it to hard tech companies, also like Ursa, right? And that seems to start to become a positive flywheel effect that I think is really going to help grow the advanced manufacturing sector uh in the US uh somewhat significantly over the next uh few years.

59:56 Uh it really kind of pops into my mind as I hear you uh as I hear you talking as I hear you um really shows the advancements that you guys have built really with a really small team comparatively to other companies. Yeah, absolutely. Um so we can track back in here on our slice. Um And Oop. So, we'll go back to desktop. So, this is um the build that we just sliced. So, it's pins and then we have sample one and sample two.

60:34 >> Yep. And then this is just a view it's a hatch view of the parts. And so, what I'm going to do is walk through the time stamp of the part. So, it's going to print So, this is the normal hatch tile filling. So, it's going from left to right uh cuz the gas flow on this machine is from right to left. So, you can see it's filling each tile like that. And then now on the second part we asked it to fill from the outside of the part in. So, now it's going from the outside into the center of the part.

61:08 Which there it go, it did work. Interesting. So, um again, not sure if it's super useful on this part, but really cool to see this completely automated workflow. Um where we prompted Claude, we took a generic file and we were able to now slice a build and loaded on the machine. So right now the build's on the machine. I'm going to go start it here.

61:38 And I'm just going to print that same layer we just looked at. And um we'll have to we'll have to get this started and then we'll watch it uh with our video to see the same hatch pattern that we just created with Claude actually printing on the machine all in the the span of about 20 minutes here. That was the coolest demo uh we've ever done on this podcast. >> [laughter] >> But I will Yeah, hopefully hopefully everything still works, but um I think this is really cool to show our workflow as well cuz if we haven't we haven't shown it in this much detail yet. Um but we can basically go and and all I did was in our slicing software I set up our working directory.

62:26 I uh put the the CAD geometry in there. And um and then I uh clicked run on the slicer and then it sliced the build, loaded it on the machine, and then all I have to do is hop in here and hit go. So it's a little So it's a little shiny. So hopefully we'll be able to see it go here, but um we should be able to see that pattern that we just looked at.

62:58 Um in a second here. Uh while we while we wait for the system to start, um so today I'm uh I'm speaking at an event called Project Manufacturing, which is a bunch of high school students um that are building uh that are working on a project uh with uh with additive manufacturing, in that case with um with uh with with FDM printers.

63:30 What's your advice to to a student that's um that was interested in manufacturing, uh probably is nervous about the future, uh doesn't really know uh should they go into manufacturing stuff or engineering, uh into a trade? What's your advice to uh to folks like that? Man, that's a really good question. So, uh it's printing right here right now. Here's the first part and it's going to print the second part right here. So, sorry for the view. It's a little shiny on the first layer.

64:02 Um Uh so, I think manufacturing is a super exciting field to be in because, you know, one of the things I hinted at before was there's a lot of challenge around um the digital world interacting with the real world. So, so I think there's a really good option both on the technician and skilled um like skilled blue collar work.

64:33 >> Yeah. Trades work, exact. So, I think there's a really good option in the skilled trades work as well as on the the manufacturing engineering side that enables um manufacturing to just scale out scale in a completely uh unthinkable way over the next few years. So, you know, I would recommend that that people find something that they're passionate about building in the manufacturing space and doesn't matter if it's a a trade or if it's on the engineering side.

65:08 Um as long as um you know, you're looking at these new technologies and you're incorporating them and seeing how things are changing. I think there's going to be a lot of opportunity uh on both sides. I think I think I think I'm I'm kind of an optimist when it comes to uh AI and that I think it's going to enable a lot more scale. And it's just like all you know, you look at all the new technologies, there's always been a uh a lot of negative press around any revolutionary technology, but every time we've had one, it's only grown everything so large that there's still opportunities for um amazing careers and you know, a huge wide range of things. So, I think that uh manufacturing is to me the most exciting field to be in right now, and I think it's going to be even more exciting over the next few years as things continue to accelerate.

66:07 Yeah, I totally agree. And it's actually there is if if you think about the role of AI, it will also make it'll make junior engineers senior engineers way faster, right? Uh cuz you don't have to go through You still of course have to go through a learning curve, but you can accelerate that very significantly. You do have uh a a buddy on your side uh that can that can guide you in a in a way uh whereas if you start uh as a as a manufacturing engineer, for example, right? You're going to you're going to have to learn everything the hard way, at least the last the last decade.

66:48 Um and I think that's going to that's going to let enable a lot of uh a lot of new engineers to get uh into a seasoned mode way faster, and that's where the fun really really starts, right? Where you can you can do work like you're doing, Tom, where um you can uh also spend some time on development and really uh trying to push the limits of a certain manufacturing technology and I think that's where it becomes very very rewarding uh really quick.

67:15 Yeah, for me personally, you know, as uh as the director running the site, you know, there's a lot of high-level tasks that I have to do. So, if I didn't have this tool, I might not get to do a lot of this development that I'm doing because I have a lot of other things that consume my time. So, this tool has kind of given me the ability to do some of this fun dev work or this exploration work that I'm probably wouldn't have had time to do or certainly not at the scale or at least that I'm doing it now.

67:46 And and so, it's just really an unlocker at every level of, you know, automating tasks that we don't want to do or giving you the ability to do a lot of this uh data interoperability between different mediums, between Excel and code and Word and even like academic academic papers. Like, you can you can mix all these mediums really easily now. And that allows you to spend like a lot of time like doing the core engineering and the the the real um fun work, I'll call it. Yeah. Yeah, it takes away the annoying part uh and and gives you exactly the fun of of engineering.

68:29 Well, Tom, this was, like I said, such an awesome uh conversation, such an awesome demo. Um I know you put uh some work into that as well because of course, if we're to uh print something, it's usually uh not something you can easily show. So, I do appreciate you uh really giving us a a peek into the development work uh at Ursa, how you guys are pushing AI, you're pushing additive manufacturing, uh you're pushing your engine programs really as a leader in that space. And uh through your openness and your willingness to share your experiences, I really do believe that you're doing the industry a great favor and you're helping the whole industry to advance and get to the next to the next stage. So, thank you very much for for talking to me today.

69:19 Yeah, and thank you so much for having me on and you know, we're really excited about the partnership that we have with EOS and we really appreciate the support, you know. A lot of this is enabled because of EOS's openness and willingness to accept new technology, you know, uh uh we were the first company that to work with yourselves to enable this new workflow and um you know, your whole team has been nothing but supportive of us. So, you know, we really appreciate the collaboration and it's really kind of given me the last year a new vigor in additive manufacturing to kind of build what you know, hopefully is the the next generation of additive, which is deploying these really advanced new software tools to enable the next level of scale and additive.

70:13 Yeah. No, you know, partnership makes makes this this industry grow and yeah, I think between our companies but also a lot of other partnerships out there are really making that making that visible. So, excited to see what comes out of Orison next. I'm sure we'll have you back in hopefully less than 12 months because there's going to be a lot a lot on the news. So, I'm excited to see what what else comes out of your your guys's organization.

70:43 And also thank you to our listeners. Thanks for tuning in. If you liked this episode, share it with a friend who needs to to deploy AI in a way that the Thomas doing it. Um a lot of efficiency gains out there to to leverage on, so uh, get get on it. This was the Out of Stack podcast. I'm your host Fabian Allafords, and I will [music] see you next time. >> [music] [music] [music] [music] [music]