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.

Waiting months for “the next big CAD release” feels out of step with how fast AI is evolving, so we sit down with Darren Henry (SVP, PTC) and Cody Armstrong (Senior Director, Onshape AI Innovation) to explain what Onshape Labs is really designed to do. It is a broad, user-accessible early access program for what’s next in Onshape, where experimental features can ship faster than the normal three-week release cadence, get real feedback early, and then either graduate into the production cloud CAD product or get shut down quickly.

We also unpack the first Labs projects and why they matter to working engineers. The Omniverse Publisher, available via the Onshape App Store, focuses on robotics by moving CAD assemblies into NVIDIA Omniverse and Isaac Sim with far less friction. Because Onshape mates capture mechanical intent and degrees of freedom, teams can define joints and physical properties like stiffness and damping, then run kinematics and dynamics simulations in Isaac. The result is a tighter CAD-to-simulation loop for robotics design, physical AI workflows, and faster iteration between design changes and realistic motion behavior.

Then we go deep on FeatureScript MCP, which turns “custom features” into something far more accessible. Instead of relying on fragile text-to-CAD outputs, the approach is text-to-code-to-CAD: an LLM generates FeatureScript, inserts it, tests it, fixes errors, and iterates until it works. The best part for teams is that the LLM cost is mostly upfront, while the finished custom feature runs natively and can be shared across the organization without ongoing token usage. We cover demos, token realities, and why persistent memory could eventually capture company standards and tribal knowledge for the next generation of engineers.

“Make me a car” is an impressive demo until you ask where the braking hydraulics, controls, and safety logic went.

We let Viral Shah, CEO and founder of JuliaHub, and Chris Rackauckas, tell us about physical AI and its use for systems analysis.

Its a different AI story than the popular one: physical AI for engineers, where models must respect governing equations, compile, and validate against known test cases. Along the way we unpack why Julia, a programming language, was created and how it led to Dyad. Hint: to do systems analyses. Any system. Also how open source shaped its growth, and why that foundation matters when you want AI to do more than autocomplete code.

We then get concrete with Dyad, JuliaHub’s domain-specific language for systems modeling and multiphysics simulation. Rather than building another CAD tool, Dyad focuses on function over form, the system and subsystem level where real products live. That unlocks fast iteration in the engineering V-model: requirements, architecture, integration, and ultimately digital twin workflows, without forcing every engineer to become a full-time programmer.

The highlight is a demo where an agent ingests NASA HL-20 lifting body documents and aerodynamic data, generates a working systems model, runs a documented pitch-pulse test, and produces plots you can compare to the original validation figures. We also talk about the trust problem with AI and why physics-aware compilers, transparent artifacts, and test cases change the conversation from “wow” to “verify.” If you care about engineering simulation, systems engineering, agentic AI, and digital twins, subscribe, share this with a colleague, and leave a review with the tool you want AI to tackle next.