AI in manufacturing has come up on Built This Week through three founders who build for people making physical things, plus our own 3D printers. Sohrab Haghighat of Hestus adds autocomplete to CAD. Hardik Kabaria of Vinci trained a physics model so engineers get thermal and stress answers while a design is still changing. Wiley Jones, a former hardware engineer, built Doss to run operations for companies that make and ship products. None of them pitch AI that replaces the engineer. Each one removes a slow step between a sketch and a product on a shelf.
There's a significant gap between here's a flashy three d model versus a three d model that is ready for manufacturing, and that is what I call it.
— Sohrab Haghighat, Episode 53
How companies we talked to use AI in manufacturing
Hestus: autocomplete inside the CAD you already use
Hestus runs on top of Autodesk Fusion, with SolidWorks due a week or two after the episode and Onshape, NX and CATIA planned later. Draw a rectangle and it offers center lines as a red overlay you accept in one click. Add one mounting hole to a backplate and you can cycle to a suggestion that patterns the rest with the symmetry constraints already applied. (Episode 53, from 02:57)
Sohrab won't ask engineers to switch tools, because companies keep years of designs in the CAD they already own. He said users have come from industrial valve design, drones, diesel engines and even dining chairs. Every design starts as a 2D sketch, so that is where Hestus started; 3D, assemblies and manufacturing readiness come next.
Vinci: physics answers during design, not after
Hardik said hardware teams have always checked physics with lab prototypes and specialist simulation software, and the real limit is how many people can run either one. Vinci's model answers thermal and thermo-mechanical questions fast enough to use while the design is still open, and he wants those answers in the hands of designers, not only simulation engineers. Its first customers make semiconductors and electronics, but he listed robotics, consumer electronics, data center racks and avionics as places the same heat problems appear. (Episode 50, from 06:19)
Whether you are designing bottle caps or semiconductors or electronics, racks in the data center, you care about its physical performance.
— Hardik Kabaria, Episode 50
The chip-specific side of Vinci, including its warpage demo, is on our AI in semiconductors page.
Doss: operations software for physical-product companies
Wiley started Doss after years of making hardware, where enterprise software kept forcing the business to run the way the software wanted. Doss packages procurement, inventory and order management on a data model each company can extend; in the demo, a coffee company tracked 25 variants of a single grind. An agent built into the platform answers questions about the business by reading the system's own workflows. His target is physical-product companies with about $20 million to a few hundred million in revenue. (Episode 44, from 02:04)
What we built
We both own a Bambu Lab X1C, and our own hardware building happens on those printers.
- 3D-printed Built This Week keychains: Sam found a customizable keychain on MakerWorld, changed the text, generated the file in about a minute and sent it to the printer. He described a home printer as having your own little factory that can make anything that fits inside it. Between us we've printed fidget tools, ladle holders, cable organizers, a phone holder and a pencil holder.
- AI product images for a 3D-printed hoodie charm store: Sam and his 11-year-old daughter print hoodie charms to sell for charity, then use Nano Banana to turn photos of the printed charms into product shots. The selling side is on our AI in retail page.
Neither build uses AI to design the object itself. That is the gap Sam wants closed:
But having an AI tool that could, like, easily customize these complex files so I could take something that already existed and and tweak it the way I I wanted would be, I think, an exciting application for AI and three d printing.
— Sam Nadler, Episode 13
What's still hard
- Pretty models aren't parts. A 3D model that looks right is not one a shop can build. Hestus is working toward manufacturing readiness one layer at a time: sketch, 3D, assemblies, then manufacturing.
- LLM progress doesn't carry over. Sohrab said physical design needs a custom model, so a better LLM mostly means his team codes faster. Vinci trained its own model from scratch and so far covers thermal and thermo-mechanical physics; customers want more.
- Specialists are the bottleneck. Hardik said physics work is capped by how many experts and experiments a company has, not by compute. Design, performance checks and manufacturability often sit with different people, sometimes at different companies.
- Old systems stick. Wiley said enterprise systems become hard to separate from running the company. Only about 5 to 10% of the market looks to switch in a given year, usually after growth, a new product line or a sale.
- Printers need care. Jordan said our printers work well but still need a lot of maintenance and upkeep, and he expects that to improve.
Episodes on AI in manufacturing
- Episode 12: Recipes from Receipts: product images for 3D-printed hoodie charms (September 2025).
- Episode 13: We 3D Printed Our Podcast: keychains on a Bambu Lab X1C (September 2025).
- Episode 44: This AI Platform Lets Anyone Manage Operations with Wiley Jones of Doss (May 2026).
- Episode 50: Physics Simulation for Chip Design with Hardik Kabaria of Vinci (July 2026).
- Episode 53: Autocomplete for Hardware Design with Sohrab Haghighat of Hestus (October 2026).
FAQ
How is AI used in manufacturing?
The hardware founders we hosted use it at different stages: Hestus suggests the next CAD step inside Autodesk Fusion, Vinci's physics model answers thermal and stress questions during design, and Doss runs procurement, inventory and orders for companies that make physical products.
Can AI design parts that are ready for manufacturing?
Not yet, according to Hestus CEO Sohrab Haghighat. Today Hestus saves clicks in 2D sketches. Its long-term goal is to take engineers from concept to a design ready for the manufacturing method they have in mind, built up step by step through 3D and assemblies.
Do better LLMs make AI for hardware design better?
Not directly, Sohrab said. Physical design needs a custom model built from the ground up, so better LLMs mostly help his team write code faster. Vinci also trained its own half-billion-parameter physics model rather than relying on a language model.
What does AI change for hardware engineers?
Vinci CEO Hardik Kabaria says physics analysis has been limited by how many specialists and lab experiments a company can afford. His goal is for any designer to ask a physics question and get an answer in minutes or seconds instead of waiting a day.