0:00 All right. Good morning. It’s so exciting to be here. You never expect engineers to say good morning back. It’s good morning. There we go. I’ve had enough of that cold brew that I’m, you know, I’m really feeling like it’s a good morning. So, this is my fourth CDFAM and I’m going to talk about ‘Text to Spaceship’. How we can accelerate our understanding and exploration of the universe using AI.

0:23 And if you’ve seen this before on YouTube or something, I’ve got a lot of new slides and a lot of exciting new videos to show. But I am going to start as I did before with structures because structures are something that people can really understand and get their minds around. They’re so physical. But I’m going to extend it to the whole spaceship or the whole spacecraft. And that’s the Text to Spaceship Vision.

1:14 And really since I first started talking about this two years ago, it’s become much more than a vision. It’s clearly becoming reality. Because if you think about how NASA missions are built, they start with a concise text description of what they have to do, what samples they have to return from what moon, or how many exoplanets they have to find life on around other stars. And then after years and decades of work and the work of hundreds or thousands of people, you’ve finally built a mission, something that you know you can manufacture and you know is going to work.

1:51 And if you think about what those people are doing for all those years, they’re mostly working on a computer these days, right? And I think that anything that a person can do with on a computer is really up for grabs in the age of AI. And if if people think there’s things that people can do on computers that AI will not be able to do, I’d be interested to come here afterwards.

2:09 So, I really do think this massive acceleration in how we build missions is truly going to happen. And you know, we’re not yet building this like, you know, generation ship that you see here that’s really just, you know, hallucinated pixels, but we can do a cube set. And you’re going to see that later in the presentation. So ‘Text to X’ has really gone from a vision to a consensus of how things are going to play out, at least with a lot of the people in this room.

2:40 And you’ve all heard that software is eating the world, right? That’s been around for a while. And it’s clear now that AI is eating software. And hardware is next. And just look at this trajectory. In 2023, we started with evolved structures. Chat GPT4 had just been released and AI for hardware was completely a novelty. We were able to make high performance individual parts using some manual CAD input, but mostly, you know, automated.

3:10 And that’s where we first, you know, came up with this bracket. And then at CDFAM in 2024, I presented the Text to Spaceship vision for the first time. The first reasoning model had been released and Text to Structure was something that we brought to the audience. You could on your mobile phone, you could actually design a structure at that CDFAM. And at that time, text to spaceship was pitched as a vision and it actually had only been funded a couple days before the keynote.

3:38 And there’s Text to Structure. But now it’s 2026 and it’s clear that agentic engineering is really taking over software and even changing who can be a software engineer. So ask a question as I did when these new ones came out. Who uses Claude code or Codex in their in their work? WOW. Okay. So so this crowd knows knows what’s happening. It’s interesting to look back when I asked about 01 and chat GPT4 probably only like 10 15% of the audience at those times had used those tools.

4:10 So the uptake is clearly accelerating. And think about where we’re going to be 3 years from now if we’ve come that far in 3 years. But you know the frontier of AI is jagged, right? It’s good at some things and bad at others. So, it’s good at language, it’s good at coding, it’s good at 2D images and kind of bad at 3D stuff, especially parametric CAD. And there are reasons for that, right?

4:35 The CAE tools, especially the traditional ones, you know, they take years to learn to use and find all the menu items and they live on desktops and they’re hard to get it into the agentic harness. Data is locked behind corporate firewalls and hard to get enough volume of. Simulation is hard just fundamentally like simulating something as basic as like friction across all regimes is actually really challenging.

4:59 And then there’s the bits to atoms gap. You can just run a program and know if the output is right, but how do you know if your telescope is going to work or your electronics box? You have to build it. So that leads to slow iteration. And I made this radar chart at the Text to Spaceship symposium 6 months ago. But just in that six months between then and now, it’s gotten less jagged.

5:24 You know, Parametric CAD, Claude has an official plugin to a major CAE platform, right? You can you can use Claude with Fusion 360, and you know, it’s really gotten pretty good at Parametric CAD. It’s gotten better at physics. Meshes have gotten almost professional quality. And you’re seeing a lot of these issues be addressed. So, the CAE tools, the major vendors are adding better API integrations. And then the folks here are making CAE tools that are AI native and data and simulation can a lot of times be overcome with the tacet knowledge we have from decades or centuries of building things right so you don’t need to simulate whether a linear bearing is going to jam right you can just use the you know standard that we have the LD ratio should be 1.5 right so you can use these rules of thumb to not have to necessarily do all the complex simulations and it’s very important that engineering organizations write those down and the Bits to Atoms gap, I think that’s one one one one of the biggest opportunities there’s still hard time getting things manufactured in an automated way and it’s really been interesting to see atomic machines particularly in yesterday’s presentation address that directly so make no mistake you all know that change is coming to hardware engineering and you know we need to message this out to the, you know, probably tens of thousands of engineers that are in this area.

6:53 I talked to a friend of mine who’s a software engineer three years ago and she said, “Oh, AI is a crappy coder. You know, it’s not going to make any difference.” And she was just looking at the point in time, not looking at the slope. And now 3 years later, it’s even changed that industry completely and who can be a developer. And we’re going to see that first in software tools.

7:09 So you see the big CAE vendors taking sort of baby steps with co-pilots and things and then of course the people in this room making new software tools lots of startups but I think a lot of you are seeing that it’s hard to get these big organizations to uptake these tools. It’s hard it’s hard to sell them and that’s because engineers aren’t used to constant change. When I started NASA we used Creo and Nastran and now you know how many years into my career we still use Creo and Nastran.

7:36 So it’s very hard to create that change. But what you’re seeing is innovative product companies that don’t need permission to innovate. They can just build things. Companies like Divergent which builds large AM structures, boom supersonic, Proteus space which builds small spacecraft and then you know Atomic Machines presented yesterday. And the mental model that I have for this is that it’s humans and agents collaborating in like a kind of a multiplayer environment.

8:14 That’s really the Text to Spaceship Vision. But I think Iron Man’s Jarvis is what people, you know, connect with much more. Really a multiplayer Jarvis where computers aren’t just calculators anymore. They are collaborators in our work. So I’m going to talk about how to build a Text to X automation. And again, I’m going to start with this EXCITE example. So, we have this football stadium size balloon that has a telescope dangling beneath it.

8:37 And that telescope can actually measure the atmospheres of exoplanets from above Antarctica. Amazing mission, but they just needed something simple. The scientists came to me, I need to attach this optical assembly to the back of the telescope. But, you know, it’s NASA. Nothing is quite simple. It had to be very stiff, had to be very light and very strong. And if you look at what an expert human comes up with, you know, they’re sketching in planes and extruding and this is probably something that like one of the Text to CAD agents might come up with too.

9:11 You know, two engineers in two days did four iterations and we could not meet the requirements. And then with evolved structures, one engineer with AI in an hour can do over a 100 iterations and come up with a design that’s lighter, stiffer, stronger, and more manufacturable. And that’s this bracket that a lot of you may have seen. And this is CNC machined. And it’s no it’s no problem.

9:33 It may look complicated and it’s curved but it can be easily CNC machined in this case by a company called Protolabs. And this has been so successful that we’ve just scaled this out across NASA and generally people should should stop making brackets. And you know we got so busy with this and Matt Matthew Vaerewyck is in the crowd somewhere. Matt and I we spent most of our time just trying to extract the requirements from the engineering team and we’re like hey we should make a checklist and standardizing our inputs is something that in retrospect is obvious but was really an unlock.

10:08 So we standardized our inputs and we standardized our outputs. People wanted manufacturable CAD and a verification report and we documented the process in text. We wrote down how you go from inputs to outputs. And there are a lot of subtleties. But now that we had this, we took the low code tool Synera And the beauty of software is that once you’ve solved it for one thing, you’ve solved it for all time to the extent that you’ve made it generic.

10:44 So with text to structure we can make something simple like the excite bracket something like this X-ray detector mount or something like this instrument mount. And one of the interesting things about this is the inputs for Text to Structure can come from a UI they can also come from an AI agent or they can of course be typed directly and this one came from an AI agent.

11:07 So that led us to the Text to X playbook and we’re now spinning up the Tex to X Factory. So first you automate your engineering process. You select your simulation and design engines. So in this case we’ve got CAD, we’ve got topology optimization and we’ve got finite element analysis and you standardize your inputs and outputs. Really important to have a defined process. You document what that process is and you create the automated workflow.

11:35 And this workflow creation, you know, used to be a big step. You know, creating these automated workflows is a lot of the work. But as you folks using Claude code know if you feed in what the inputs and outputs are and that document that can actually start writing a lot of the code for you and you also have to verify it with evals. So you have to have applications that you can test it against so you know that it works and you have confidence in the results and then finally you scale it using cloud deployment.

12:04 So you know you prototype it, you share it out to early adopters, you iterate on it, you democratize it to your community in the cloud and you sustain it and build communities around it. And this is what we call the Text to X Factory where we take engineering workflows and automate them, put them in the cloud and then we can start linking them together with agents and it becomes very very powerful.

12:25 So that was just Text to Structure relatively simple. Let’s look at something more complicated. Text to Telescope. Here we need a supervisor agent and it has to have some folks working for it. Optical design agent, an optical mount agent that encodes server knowledge of how to mount things, an analysis agent and a reporting agent to create that verification artifact. And they each have workflows and they also each have engines, right?

12:53 So underlying the optical design, you’ve got a ray trace engine. And underlying the mount, you’ve got a CAD engine. And of course, you have to have a simulation engine. And one of the beautiful things about this is if you want to replace your old CPU bound, you know, simulation engine with with a new cheaper faster, you know, GPU based engine, you can just drop that in to the extent that you standardize your interfaces, run your validation so you know that it gets results and keep moving.

13:19 So let’s look at what this is like in action. So you start with wiring up your large language model to your supervisor. You’re telling it what’s its roles and responsibilities are, what its goals are, and give it rules and boundaries. So, you’re essentially doing context engineering for each agent. You wire it up to the other agents that have the various workflows that I was talking about. You know,the CAD, the optics, the simulation, and then you publish it to your secure cloud environment.

13:47 And now you can just start talking with that supervisor agent about the kind of telescope you want to build with, you know, basic things that you need. You need the aperture, you need the focal length, field of view is important, and then you can have a back and forth conversation with that agent before you get started. And then you hit go and you’re off to the races.

14:06 So, it’s running the various agents. They’re talking back and forth to each other. And the beauty of this is you can actually see what they’re saying, right? You have a trace. You can see what they’re saying. You can see those standardized inputs that they’re passing back and forth. There were all the optical parameters and then it took those optical parameters, made CAD models, made CAD models of the mount, made CAD models of the telescope tube and then it does the simulation to make sure it’s stiff enough.

14:35 Check the factors of safety and then finally it’s going to go ahead and produce the analysis the verification report. So something that had taken months can take minutes. And you know, I’ll admit that there’s a lot of details that need to be added into this like how you, you know, what adhesives you use for the optics and how you align them. But you know, clearly this kind of technique is going to work.

14:58 I’ll give you another one. This one hasn’t been seen before. Text to bipod flexures. So bipod flexures mount a lot of our precision instruments. So if you have, you know, a mirror or some other instrument that like an optical bench that needs to mount to your spacecraft, bipod flexers provide mechanical and thermal isolation. So they’re usually like a neck down flexure and then a carbon composite tube is how we usually make them.

15:20 But getting these right because they have tight constraints can take again months. And again, we’re going to use Sinera. They helped us put this together. And we’re going to have a supervisor agent as we did before that manages the process. We have a logging computing agent I’m going to talk about in a minute. The design agent that makes the CAD, the simulation agent that does stress, does modes, and does thermal because these have to provide thermal isolation.

15:49 And you know, then you’re just going to wire them up. So the logging and computing agent, we’re doing a lot of iterations. We’re exploring the design space with this and the logging and computing agent they have found makes things more reliable. It’s kind of like like a bookkeeper. And here you’re going to say I want a flexure bipod flexure system. In this case it holds a 15 kg instrument, you know, 100 mm off of the deck.

16:16 So that sort of sets the the broad parameters of what it is that you want to build. And then it goes off and does its thing. You know, it runs the CAD agent, it builds CADs, it runs the simulation agent, it checks what the performance is, and then it iterates on each of those. And in this case, we’re using this like two-stage iteration because it has to sort of do the tubes separately.

16:39 And then finally, again, you get that verification report and that shows what iterations it did and verifies that it meets the requirements. So again, something that definitely took months. You know, I’ve built these things, worked on them particularly for the Roman Space Telescope, but they’re used on habitable worlds. They’re used on Hubble. We can vastly accelerate that. It’s very useful tool. So, this is the Habitable Worlds Observatory.

17:02 It’s a three story tall telescope that’s NASA’s next flagship, and it’s going to be able to find life around exoplanets. To build something like this with AI, we’re going to need automations for everything. For the micromedorite shields, for the solar arrays, for the high gain antennas, and we’re going to need to stitch them all together, and we’re working on that. We’re not there yet, but for a cubesat, we’re getting pretty darn close.

17:30 So, what you’re going to see here is Celedon, Da Vinci, they’re the next talk, and a paragraph describing a cubesat is pasted in, you know, basic things like what its orbit is, how much it weighs, what pointing requirements it has. And Da Vinci makes a plan and executes it and it’s going to develop all the systems model and the things a systems engineer would expect that sort of define the mission in the systems engineering set requirements block diagrams etc.

17:59 So this has been seen before but now we’re going to take it to the next level where we’re going to start building it out with actual components. So it goes out to the web and grabs a reaction wheel that meets the requirements and then it takes the parameters of that from their documents and pulls it in. But this is the next step. We’re going to actually pull in the geometry.

18:22 It’s going to go out to on shape, grab that CAD model and pull it in. And it’s going to detect what the mounting features are. And it’s also going to figure out how to place it in the systems because a lot of things are really just components placed in a system held together by structures. So it’s going to think about how to place those and it’s also going to build out the whole QAP frame because you know for basic stuff like this text to CAD just works now with the latest models and it puts them in smart places and now we’re going to go ahead and call the text to structure agent and we’re going to mount those you know it’s it’s a lot of interfaces it’s like 12 interfaces and we’re going to call the Infiniteform engine and go ahead and build an optimized structure that holds those together.

19:04 So here you have the optimized structure. You’ve put it in there and you can just continue to do that. You just grab your components and have them mounted to the structure and you can really really quickly build up a spacecraft bus. So here we’re going to complete the spacecraft, generate all the remaining geometries. Some of them are just kind of like at the volumetric level. So it goes ahead does Text to CAD lays the whole thing out and now you have a pretty good concept for a six cubesat done extremely rapidly.

19:42 So it’s clear that Text to X and Text to Spaceship is maturing extremely rapidly. So what does this agentic engineering future look like? It’s an ecosystem of cloud deployed agent accessible design tools that are be both internal and external. So something like habitable worlds might have a lead agent that calls a telescope agent. But maybe this time we want an off-the-shelf optical mount goes out to Thor Labs if they had an API and pulls that information in.

20:19 And then you’re also going to see it for more complex things. Say you need a custom battery. I mean this is a complicated mission, right? You may need you’re going to need a custom battery. You go out to the battery vendors. You speak to them in an agent to agent kind of way. Hand off the requirements. They can have a conversation. Maybe you have to wait a day and then you get the artifacts that you need to continue your design and run these massive loops very quickly.

20:37 Let me see how much time. Okay, I’ll skip this. You guys know what agents do. But I do want to show this. So Text to Spaceship in action. We were given, you know, Matt and I were given the opportunity to fly something on one of these scientific balloons, but the caveat was from the kickoff meeting, we had 12 weeks to bolt it on, which is pretty darn fast for NASA.

21:01 So, we used AI to figure out what we should even build in that time frame. We used it to select the components, to write the software, to design the structure, and it flew successfully. And, you know, one of the things that AI recommended is that we put a selfie cam on, and we’re very glad we did. So to conclude, Text X agentic engineering for hardware is happening now and is radically changing the way that missions are developed.

21:31 Faster, better, cheaper hardware really unlocks new capabilities in our world. And I think it’s a good time for us to think about what we want AI to do for us. Do we just want more stuff or do we want more human connection? Do we want robots endlessly battling each other, or do we want to go out to the stars together? For me, I want to understand and explore the universe with AI, driven not by optimization, which is what computers are good at, but by the human values of curiosity and connection. And it’s the people in this room that are going to make this happen.