0:00 These are some other examples that I just wanted to show you. So, we’re doing large assemblies that are completely AI-driven like the suspension on trucks. We’re doing heat sinks where we’re calculating behind the scenes the thermal properties as well. We do we we have a shell and tube heat exchanger. We’re not using we’re not using lattice structures in ours, but it did create this is a over a 500 component shell and tube heat exchanger.

0:24 The water tower is probably the largest one we’ve done. We have a engineer on staff that did a water park and it will it can do conceptual tools as well as finished goods. Deterministic, it generates the tool. The tool then has a deterministic result every time you use it. Recorded at the CDFAM Computational Design Symposium in Tokyo, October 2026. Learn more at CDFAM.com. Hi everyone. My name’s Darren Henry.

0:58 I’m one of the executives at PTC. If you’re not familiar with PTC, we’re a company that makes engineering software. We make a software called Creo. It’s a CAD package, one of the first parametric CAD packages available on the market. We have one called Onshape. It’s rather revolutionary because it’s a cloud-native CAD system. It works in a browser with zero software installed. What I’d like to talk about is autonomous design and AI-driven design and how we’re bringing that into mainstream CAD tools.

1:35 I live in a world of NURBS, B-reps, and parametrics. And I think most of the engineering that’s done and most of the products that are manufactured are in that format. So, I’d like to talk a little bit about what we’re doing to bring AI-driven autonomous design into that environment and and what our philosophy is as we look at these tool sets. Right now, AI delivering impressive results with CAD.

2:03 You may not be familiar with Onshape. Here it is running in a browser, but you’re going to see tools like this where you’ll be able to prompt AI to either query or look for information or do tasks within the CAD application. Here I’m asking it how much would it we save in weight by switching material of the side panels of this trailer, and it quickly iterates and looks at the the geometry, the context of my design, and in this case it says that switching from steel to aluminum will save us 266 lb, and then we could tell it go ahead and switch those material properties, export those aluminum panels, maybe maybe you could even have it tied to other systems and send those files to a vendor for manufacturing.

2:52 The key is that it’s using natural language to do commands within the software in the context of your design. So, it acts as either an advisor or an assistant. Okay? This is another example of a tool. This this one’s in early visibility for us right within the CAD system. You know, instead of using a ray tracing rendering package where you’re sitting up you’re setting up the environments and the light sources, you’re just giving it sort of a diffusion prompt where you’re saying, “Hey, take this model in the current orientation, put it on an engineer’s workspace.” And it’s using AI to go ahead and create this.

3:29 Now, you could do this outside with a with a screenshot, but what’s nice is when it’s tracked within the CAD software, the engineer then has that history with them, and they can always refer to it, make a design change, and then re-render it in the same environment very easy. It’s more predictable. So, you’re going to see tools like that. That’s going to be table stakes for CAD applications.

3:55 But what everybody’s asking for or what gets a lot of hype in the CAD market right now is the idea of text to CAD. It’s a person that doesn’t have maybe CAD experience or doesn’t want to put the effort in to building something step-by-step with CAD. So, they just want a prompt and have the geometry created. And this is actually sort of an interesting example in that we did this 2 days ago with Astra, the latest AI model, and we’re using Onshape here, but we we did two AI-driven designs with the exact same prompt, exact same setup simultaneously.

4:32 And what’s interesting here, let me read the prompt for you. Make a fixture to hold workpieces when welding two pipe pieces at 90°. Make it so the pipe diameters can vary between 1-in and 3-in diameters. The entire fixture should be able to be bolted to a table or workspace, right? So, this is a typical fixture that an engineer might have to create. Unfortunately, while the results are compelling and these designs would work, it’s it’s a little bit scary that the same prompt yielded such different results.

5:07 And in fact, if you look at this, one is 35 parts, another is 49 parts. But if we look closer, we realize that not only is it different part counts and slightly different designs, the same prompt yielded different CAD construction techniques, different shapes, and and different dimensions on similar parts. Now, the other thing is so this is not deterministic. And what’s key for us and as we talk to different manufacturers, while this is really good for conceptual design, if you were looking to manufacture and improve this, you don’t know which design’s best without some type of validation.

5:46 And you realize if you ask the same prompt again, you might get a a a drastically varying result. The other problem is that in many cases they’re either oversimplified and what we call brittle models. And when I say over simplified, you can see that as I zoom in different levels of detail, probably not what you’d expect from a human CAD operator. And they’re not easy to edit.

6:17 They’re built in a different way than a human would build it. And if a human wanted to make a change, chances are they’d have to prompt re-prompt with an LLM instead of just making the change in the CAD tool. Now, we think the models are going to get better and better with time, and this is going going to improve for all CAD vendors. But right now, we feel that it validates something very important to us, which is humans need to be in the loop when you’re talking about AI-driven design, at least in the mechanical sense.

6:51 So, we’ve actually developed internally, and these are our beliefs, three rules. And the first one is humans need to be in the loop for verification and making intelligent edits. Now, the intelligent edits is is probably the new piece of my talk. I think we’re all all the different presentations have mentioned human in the loop for validation. But it’s important. You want to interrogate the model, look at different things, check fit and function, maybe bring in other geometry and check fit and function with with with your new design.

7:21 So, it’s it’s good news for us where we believe the CAD applications that we’re using today on the market will be the meeting place between engineers and AI. But we also want it to be deterministic. We want to get similar results every time you’re actually prompting and working with AI. And that’s a challenge. And we want the design intent to be preserved. Meaning, if I change those pipe sizes, we’d want that fixture to accommodate.

7:46 Or if I asked it quickly, let’s make it make it I don’t need a three 3-in pipe or a a 30-mm pipe, I want to go up to a 50-mm pipe. So, redo this design. You want a deterministic where the design looks very similar, it just re-sizes intelligently. So, we feel it’s very important to have AI build the models where the engineers work. And Onshape, the CAD application that we’re building or that we’ve built that’s cloud-native, it runs in the cloud, it’s completely scalable on AWS infrastructure.

8:21 It actually can satisfy a lot of these rules with something called FeatureScript. And let me explain that. So, Onshape is sort of a next-generation CAD application. And it’s very fortunate that when we launched this, we decided to expose the language that our engineers use to actually program the features. So, if you’re familiar with CAD, there’s a sketch command, an extrude command, a fillet command, a shell command.

8:53 That’s language that the the vendors, developers actually write on top of what they call a geometry kernel. I won’t bore you with the details, but Onshape’s unique in that we decided to expose that to the public. And our intention at the time was that the average user with a little programming experience could then customize the CAD solution to meet their needs. So, a good example is we have a company that makes cabinetry.

9:23 And they use that programming language to build special at you know, if you go to IKEA, you have special fasteners like barrel barrel fasteners and so forth. So, they made special features so they could add those very quickly to their designs. It did require programming expertise, so it wasn’t something the mainstream adopted, but very but a few people did and they were able to really customize the CAD package.

9:49 It does everything we want. It captures design intent. It yields deterministic results. It meets you in where you work in the CAD application, and it actually builds a little interface for you. And the beauty of this is AI loves it. Large language models really appreciate it. And what we did is we actually stuck it and connected it through an MCP server so that LLMs could actually work in FeatureScript.

10:14 It not only programmed it, it could debug FeatureScript, it could check the results, and build you a custom engineering tool to meet your needs. You’re customizing CAD to suit your products and processes in-house. This is critical. So, we call it text not text to CAD, we call it text to code to CAD because you’re texting an LLM, it’s generating code for you that has deterministic results inside the Onshape application.

10:49 Not only does it iterate to get it correct, but it remembers each project, so it has a memory and gets better and better while it works. Let me show you how it works. So, in this this is going to be a simple example. In this example, I have a PDF of a bushing. And I’d like to It’s something that my company has to make over and over again.

11:13 So, I go into Claude, and this is Opus 5 at the time, and the top line is my prompt where I upload the PDF, and I say I want a custom feature to make this bushing. And Claude calculates. And you probably can’t see it, but it’s doing a lot of discussion with me. I’m not talking to it. This is one prompt where Claude came back, and after 36 minutes, it generated all this geometry.

11:35 And this, you might say, “Okay, that’s text to CAD. You You text it, and you got a CAD.” But it actually wrote code behind the scenes. And then, this is its sandbox, then in my document, and this runs in a browser, it actually improved that code to have an interface and become a an engineering tool that has design intelligence around the bushing. And now I have a new feature that I can generate the bushing, and it’s not only one bushing, I can generate all the bushings that were in that PDF.

12:09 And immediately, because it’s in the cloud, like you’re seeing here, we can share it like a Google Google Doc throughout our entire organization. And now my company has a unique CAD advantage in making bushings. It learned from my history that I like to change the color, and I also like the blue callout on the top with some important information. It put that in automatically because of previous project that I did.

12:31 So, it’s quite interesting that we were able to build a custom flange oil-free bushing feature in Onshape with one prompt and 36 minutes of cloud’s time. Let me show you another example. This one also is a component, but this will show you the benefit of how having this inside a production-level CAD system can yield some nice results. So, this is a roller chain sprocket. If you know anything about roller chain sprockets, the tooth profiles are extremely difficult, right?

13:03 But you can implicitly develop them with an equation. So, we loaded up some information. It went to the web and and grabbed some information, and now we have a custom tool for building sprockets. And look at all the variations that it came up with for us. We can do hubs. We can change the fastener from a keyway to a set screw or or vice versa. We can have multiple strands of the sprocket.

13:28 You notice the dimensions. This is GD&T information. We call this model-based definition, so that it stores all that information in a nice table automatically because it’s utilizing the production CAD system, right? It’s not a third-party. It’s not It’s not a special environment for AI. It’s you doing it inside the CAD system. And you can see the cross highlighting as I look at that. So, a very powerful sprocket configurator that I can share, and I can include in human-made machines because it’s built natively in the system that I work in.

14:07 I’m going to continue with some examples. It’s not only components, so I showed you components like a bushing or a sprocket, but I I to broaden your mind. Here’s a half-tone sphere pattern. This is a a really neat site called Book of Shapes, where you can look at different patterns and be inspired by that. And I wanted to create a feature that I could apply to any part.

14:28 Maybe I’m a speaker manufacturer, or I’m making smoke detectors, or I do shower heads. So, I can build a custom feature based off that image. It was an SVG file that I uploaded into Claude and a prompt, and it built me a custom feature that’s extremely difficult for you to make manually, but AI was able to make it very quickly. And in fact, there’s a rectangular pattern, but I can pick a staggered pattern, which also looks neat.

14:55 I can have it diffuse over a radius or fill the whole thing. And again, it’s building the code behind this feature, so it’s a reusable engineering tool. I used Claude tokens to build the geometry, but now that it’s a tool, it’s not using Claude tokens anymore. It’s just using Onshape. This is a special one called a phyllotaxis that I really like, but unfortunately I’m I’m not going to go back, but unfortunately with the phyllotaxis, it’s not symmetric.

15:28 So, instead of me trying to figure out how to make this complex pattern symmetric, I simply asked Claude to do it as a second prompt. It built me one that is symmetric, and I can make changes with even an offset in the center. So, you can get rather complex features. This isn’t building the It’s not building the square chassis, it’s just building the whole pattern that I’m applying to a surface.

15:53 I hope that makes sense to everybody. Can someone clap or or nod if it does? Thank you. Okay. I I didn’t mean to ask for applause, I just wanted to know if you understood. All right, this one this one Okay, this is an interesting one. Let me set this up. I have a fantasy about this one. So, what I’m saying is you can now customize your CAD to meet your needs.

16:14 Before, an engineering company would have intellectual property around the models and the drawings and their products. Now we’re saying you can have intellectual property on the engineering tools that you built yourself with AI. It’s critical. So here’s my fantasy on this one. A young engineer inherits his father’s business. His father had a a business making wire works, wire goods. And he’s trying to improve not only the markets he can penetrate with his wire goods, but the efficiency of the company.

16:49 So he gets Onshape and Claude and our FeatureScript MCP, which is free, and this is what he does. He makes a new tool where he can just click on a rectangle and have a grill rack, and then he can use his new tool to change the orientation and spacing and sizing for any new customer that makes grills. And it makes the cut list because it works in the CAD system, makes the cut list.

17:16 But then he says, “You know what? I’m going bigger than grill racks. I’m going to do sink racks as well.” So he builds a tool again, and this is maybe an hour’s worth of prompting, where he can actually make custom sink racks very quickly. And you can see he can resize it, he can drag around the drain. Does he stop there? No, he goes right to industrial racks.

17:39 And here you can see he built a nice tool with a zigzag and he can change the spacing, but more importantly, he can actually go ahead and build these wire racks instantly, or these shelves instantly, with any size rectangle that a customer gives them, he just clicks. So he takes his dad’s intellectual property, how they do things in-house, he programmatizes it, and now he has a tool that’s incredible.

18:05 Does he stop there? No, he’s on a he he’s on a tear. He actually starts looking at wire baskets. So he makes a new tool with a prompt that does wire baskets. Here’s one that’s stackable. So, he picks a solid and a face to open up. It makes the ring in the hoop automatically. I build all these features myself. I did it in less than a day.

18:32 And what’s beauty the beautiful thing about this basket is you can actually have tables that show you not only the bend information and the cut information for all the not only the cut information for all the wire strands, but the bend information for all the wire strands as well. So, I think it’s revolutionary. I think we’re on a new era of customizing CAD. And if you can define any geometry with an equation, you should be able to programmatize it with with this tool.

19:02 And it’s called Feature Script MCP. These are some other examples that I just wanted to show you. So, we’re doing large assemblies that are completely AI-driven like the suspension on trucks. We’re doing heat sinks where we’re calculating behind the scenes the thermal properties as well. We do We We have a shell and tube heat exchanger. We’re not using we’re not using lattice structures in ours, but it did create This is a over a 500 component shell and tube heat exchanger.

19:29 The water tower is probably the largest one we’ve done. We have a engineer on staff that did a water park. And it will it can do conceptual tools as well as finished goods. Deterministic. It generates the tool. The tool then has a deterministic result every time you use it. This is also a a key part. It doesn’t always have to create geometry. One of our engineers has loaded up the FSAE rule book.

19:59 This is a college competition. It’s about a 500-page rule book. And he asked Claude, I think he used Codex actually, not Claude. So, he can work with any LLM. He asked it to create a number of custom features that check the design against the rule book. And as you can see, it made a table. He has six failures, three warnings, and 88 passes. All with custom features that he built in a day based off the PDF FSAE rulebook.

20:35 So it doesn’t only have to create geometry for you, it can do calculations. It can actually do measurement tools. We have one that actually will fill certain containers with different volumes so you can see the weight, the center of gravity shift. We have buoyancy calculators that we build as well. This is my summary slide. So if I asked you to remember one thing, two things, remember Onshape.

21:04 If you don’t have access to it or you don’t know what it is, there’s free versions that you can play with. And MCP is free as well. You would just need to bring your own LLM. But this is what I really want you to remember. Now AI can build production level CAD tools that are customized and tailored to your processes and your products. You can turn prompts into reusable intelligent CAD features.

21:31 It captures design intent and it’s deterministic and that’s key for us. And the engineers are always in control. It’s it does the work where they work. You can check it using all the tools that you’re used to check your own work. And it should be a familiar interface with you because it it’s a CAD tool that you may already use. That’s my presentation and I appreciate your time. Recorded at the CDFAM Computational Design Symposium in Tokyo, October 2026. Learn more at CDFAM.com.