Most discussions about AI + CAD focus on the wrong layer.

Developers often compare:

model performance

geometry generation capability

plugins and toolchains

But in real engineering workflows, the bottleneck is rarely generation .

It’s this:

Can the output actually be executed?

Is it consistent?

Is it manufacturable?

The Idea

Instead of asking GPT to “generate a model”, I constrained it to do only one thing:

Convert a single natural language input into a runnable CAD build script

Example

Input

outer diameter 120, inner hole 50, thickness 15, 6 M8 holes, PCD 90

Enter fullscreen mode

Exit fullscreen mode

Output (FreeCAD Python)

import FreeCAD , Part , math

doc = FreeCAD . newDocument ()

OD = 120 ; ID = 50 ; T = 15 ; HC = 6 ; HD = 8 ; PCD = 90

outer = Part . makeCylinder ( OD / 2 , T ) inner = Part . makeCylinder ( ID / 2 , T ) body = outer . cut ( inner )

r = PCD / 2 holes = [] for i in range ( HC ): a = 2 * math . pi * i / HC x , y = r * math . cos ( a ), r * math . sin ( a ) h = Part .…