Peggy Xia, cofounder and CEO of gNucleus, appears on the Future of Design and Engineering Software podcast to cast light on what gNucleus is — and isn’t.
I give up on text-to-CAD again, the way I have given up on it several times before, but the last time felt different. I had selected a text-to-CAD tool at random. It didn’t go well, as I have chronicled, but it led to an exchange with Peggy Xia, co-founder of gNucleus, the company that had the misfortune of being selected for my latest failed attempt at text-to-CAD.
Normally, exchanges between journalists and their victims don’t go well, either. But there was Peggy, to her credit and my relief, gracious in defeat — and eager to dissuade me from the notion that gNucleus was a text-to-CAD company.
“gNucleus is a model company, not a text-to-CAD company,” she said. I was confused. In my world, a “model” is a 3D CAD model. In Peggy’s AI world, “model” refers to an AI model, such as OpenAI’s frontier model, GPT or Anthropic’s Claude.
The text-to-CAD tool, which I had found underwhelming, was not the main product but simply a public-facing tool for people to try out, a flame to draw moths.
[In all fairness, gNucleus was able to add more data to its CAD generation model after I uncovered the flaw and the second attempt improved the shape of the bracket.]
Meet Peggy Xia
I invited Peggy Xia onto my podcast, Future of Design and Engineering Software, to find out what the company is actually building, which indeed was considerably more than modeling a simple sheet metal part I had tested it on.
You can watch the whole interview here and listen to it here.
Peggy has a deep background in CAD. She started her career in Shanghai with Siemens, where she worked on NX’s core part modeling functions, specifically the Extrude and Revolve functions. She worked on Synchronous Technology, a technology so advanced it could be considered early AI. Its direct-modeling approach offered an alternative to the Jenga tower of history-based, feature-tree solid modeling. Then she spent roughly two years at General Motors Account at HPE working on their CAD/PLM systems. From there, she moved to Hexagon offices in Southern California, working on pre-processing, meshing and generative design.
It was then that the newly emerging field of machine learning caught Peggy’s eye, and her career took a sharp turn towards AI. She moved to Amazon and then to Google where she developed YouTube’s recommendation models.
“There are around two billion people a day using the recommendation model I built,” she says. The model can recommend YouTube videos based on their quality, but she can’t say much more since tech companies’ tech is closely guarded secrets.
While working on YouTube’s Recommendations team, Peggy met Mei Chen, an experienced AI researcher who had previously worked at Google Brain and later Google DeepMind, where she also contributed to Gemini’s post-training.
Despite the success and popularity of her recommendation engines, Peggy realized that guiding people to more video was not exactly a noble calling. She longed for more serious application of her skills. And what was more serious and more in need of AI than CAD?
About gNucleus
“We are still a small company, we have about 10 to 20 people,” Peggy said — split roughly evenly between AI researchers and engineers.
Peggy and Mei are betting that shape can be controlled like language. gNucleus is trying to train a foundation model that understands three-dimensional shape and manufacturing intent the way LLMs understand grammar and meaning.
Peggy described the training process in three stages: pretraining on large volumes of unlabeled geometric data, a post-training stage that uses reinforcement-learning-style scoring to reward better outputs, and a final fine-tuning stage on a customer’s own proprietary data — their actual parts, their actual manufacturing rules, their actual tolerances.
“If the model didn’t see such data enough, it will start guessing — hallucinating,” she said. “If you feed more data, then you perform better.”
A sheet-metal bracket bent into shape is a specific geometry that could fool a general-purpose demo model such as their own, Peggy says. But all that was needed was to feed it some sheet-metal parts.
Peggy tells us how the gNucleus training process is actually paying off with enterprises the company is engaging with. Their AI model gets fine-tuned on a real customer’s parts, not generic shapes.
“We can generate ninety-five percent of an electric motor already,” she says. “Every major component except the housing.
Compare that to a human engineer. It would take them weeks to make an electric motor from scratch. On simpler, more repetitive geometry — a gear, for instance — a version of their model fine-tuned for that task can produce a usable part in two to five minutes, compared with fifteen to thirty minutes for a general-purpose frontier model, such as those by OpenAI, Google and Anthropic, asked to do the same thing without specialization.
None of this is trivial development. Peggy knows enough about CAD models to understand how hard it is to turn messy real-world inputs — point clouds, meshes, photogrammetry scans — into clean, parametric CAD models that an engineer can edit.
Entertainment and gaming applications can tolerate meshes that are visually convincing but geometrically sloppy; manufacturing cannot. A part that will be machined, pressed, molded, etc., may have fine tolerances, smooth surfaces…
gNucleus publishes CAD benchmarks on Hugging Face comparing token costs and accuracy across frontier models, making the case publicly that a specialized, fine-tuned approach is both cheaper and more accurate than throwing a general-purpose foundation model at CAD geometry.
Maybe LSM, short for large shape model might be a better category for the type of AI gNucleus is creating.
“We are an engineering AI model company,” she said, drawing the comparison explicitly: “like OpenAI — we build models, and any other startups or CAD companies” can build on top of them.
Small Company, Big Ambition
The name “gNucleus” comes from “generative geometry kernel,” substituting “nucleus,” Latin for “kernel, explains Peggy.
gNucleus wants to do for design what OpenAI did for language — create a foundation-model layer that other, more consumer-facing design and CAD tools could eventually be built on top of, the way a wave of AI apps has been built on top of GPT, the way CAD apps have been built on Parasolid.
gNucleus really wasn’t just a text to CAD company or sketch to CAD company (another tool in their showcase not discussed here). The company’s mission was much bigger. It was building a CAD foundation model that could be licensed and fine-tuned for enterprises willing to feed it their own products, so it would be trained to generate other products.
gNucleus is led by an all-star cast. Along with cofounders Peggy Xia, who serves as CEO, and Mei Chen, CTO, is co-founder Roc Mao, head of engineering, with fifteen years in CAD, CAE, and CAM, including stints at Siemens and Hexagon. Eva Lu, the company’s COO, previously worked at Google Cloud, IBM, and Accenture. The company is headquartered in Sunnyvale, California, in the heart of Silicon Valley, and home of Alphabet, better known as Google.
gNucleus has raised one funding round to date, a pre-seed round led by the venture firm Baukunst, whose CAD-savvy investor Axel Bichara has led investments in SolidWorks, Onshape and is currently invested in Motif, founded by former Autodesk co-CEO, Amar Hanspal. The amount gNucleus has raised has not been publicly disclosed.
