It works by teaching an LLM a compressed DSL representation for footprints. I think this is generally a good technique if you're trying to get LLMs to work on spatial problems.

Right now I'm focused on reproducing footprint shapes, but after that I'll benchmark the correctness of the footprints using % overlap on known package descriptions, as well as a kicad_mod export

I built this because I realized I needed to use AI to help tscircuit catch up to the decades of electronic design tool progress that I'm not able to utilize (as an MIT-licensed, Typescript project)

Source code here: https://github.com/tscircuit/text-to-footprint

Feedback/critical opinions appreciated!