A lot of “AI for CAD” talk still feels like a wrapper around old tools, so we wanted to get closer to the engine room. Lewis joins us to explain how he went from years at ACIS and is now building his own CAD kernel and a browser-based solid modeler. It is wired it up to Anthropic Claude through an MCP plugin so the model can run tools on a real computer and output real CAD data.

We dig into what makes text-to-CAD genuinely useful for engineers: not just generating shapes, but producing editable STEP files, surviving fussy details like fillets and blends, and behaving less like a guesser and more like a careful intern who can learn fast.

Lewis walks us through why “train on how experts used SOLIDWORKS” is not the same as building reliable geometry, and how constrained prompts plus workflow metadata can keep AI work from falling apart as tasks grow.

The demos are insane! Not only had Lewis managed the Great Bracket Challenge (that's a joke), but also a drafting table and, hard to believe, a variable stator for a turbine. A gas turbine can be a jet engine or a power-generating device. Boom Supersonic bragged about creating one. Lewis had FeatureBoard, his CAD program, create what looks like a fairly complete CAD model from four screenshots.

We follow the thread into synthetic test data for kernel hardening, automated bug logging, design-for-manufacturing checks, technical drawings, and even the path toward auto-quoting. We also touch on robot welding simulation, interoperability with STEP AP242, and why “foundational CAD models” may be less mysterious when you view geometry as language and computation.

Your CAD tool can’t read your intent, but your fabricator still has to. That gap is where AI CAD promises a lot and often delivers noise, broken geometry, or subscription regret. We sit down with Bobby, a mechanical engineer and founder of Transfigure, to get specific about what actually works today and what still fails when you demand a real STEP file you can open and use for getting a quote or for manufacturing.

Bobby started with a blunt experiment: ask an LLM to generate a simple one-inch cube as a STEP file. The output looks believable, yet it the file is empty.

From there, we unpack why “text to CAD” and agent-driven kernel scripting can hit hard limits, especially as geometry gets more complex and as file sizes grow from kilobytes to megabytes and beyond.

Bobby explains Transfigure’s current 2D to 3D workflow, why engineering drawings, sketches, and even photos are a practical input, and how the goal is to hand you a manufacturable STEP file instead of trapping you in a feature tree. We also talk sheet metal realities like bend radius and standardized thickness, the quoting pain machine shops face when customers only have 2D, and a real-world demo flow that goes from a sketch to Fusion to CNC with partners like Tormach. We close with the business side: per-generation credit pricing, why many AI CAD tools feel “tainted,” and why a bespoke foundation model may be the only path to reliable AI-generated geometry.

A text prompt that “almost” makes a bracket is the fastest way to lose an engineer’s trust. I sit down with Peggy Xia, co-founder and CEO of gNucleus, to find out what, besides almost making a bracket, gNucleus was up to.

Peggy's background is not that of a typical AI startup. She has had a career in the real mechanics of engineering AI, CAD automation and simulation workflows with Siemens. He wrote the code for Solid Edge's Synchronous Technology. At Google, she developed YouTube's most widely used recommendation model, which emphasizes video quality.

We unpack why text-to-CAD goes wrong in ways that feel unforgivable to practitioners: the model is forced to guess when it has not seen enough domain data, and that guessing manifests as hallucinations such as misplaced fillets or incorrect features. Peggy explains the training stack in practical terms: pretraining, post-training with reinforcement-style scoring, and fine-tuning on customer-specific CAD data, so a model can become genuinely strong in a narrow domain such as automotive motors or assemblies. We also talk about multimodal AI, what “sketch to CAD” could look like, and why converting meshes or photogrammetry outputs into clean, manufacturable parametric models is still one of the hardest problems in the pipeline.

From generative design to computational design, we challenge the idea that a cool-looking shape equals an engineering solution. Accuracy, tolerances, benchmarking, token cost, and runtime matter, especially when you want production-ready results, not just visuals.

If you care about the future of CAD, CAE, and manufacturing-grade AI, this conversation will sharpen your mental model of what’s possible now and what still needs breakthroughs.