AI Can Now Turn Plain English Into Factory-Ready 3D Designs

If you've ever hired a drafter or waited weeks for a product design, this is for you. Engineers design physical things — brackets, phone cases, car parts, machine gears — using CAD software (computer-aided design, the tool that draws precise 3D shapes). Today, AI chatbots can sort of do this from a text description, but they're unreliable. They're guessing machines. And a guess that's half a millimeter off means a part that won't fit.

The research team behind ReliCAD attacked that gap directly. Instead of letting the AI jump straight from a sentence to a finished shape, their system does three things in order. First, it translates your request into a structured set of design rules — what connects to what, which edges must stay parallel, what must be built first. Second, it feeds those rules to real CAD software, which does the exact math. Third, it inspects the result, finds where the model drifted from the original request, and fixes it — then tries again until things line up.

The payoff shows in the numbers. On public test data plus the team's own editing tests, ReliCAD produced valid, buildable models 99.8% of the time, with a high shape-match score (0.8753) against what was actually intended. For comparison, earlier AI approaches frequently produced models that broke, contradicted themselves, or couldn't be manufactured at all. A near-perfect hit rate changes this from a fun demo into something a small machine shop could actually use.

There's a catch. This is a research paper, not a product you can buy today, and it was tested on standard design datasets — not the messy, one-of-a-kind requests a real customer walks in with. It also still needs a human to describe the design clearly and to sanity-check the final model before anything gets cut in metal. Still, the direction is clear: the gap between describing an object and having it engineered just got much smaller.