Pure-C trained GPT, INT8 quantized, inference in vanilla JS. No bundler, no WASM, no committed HTML/JS — everything (including the model weights) was served from a single workflow YAML.
_