Live demo
An 80 KB int8 ONNX model running on WebAssembly — the same model the CLI runs, no server and no upload. Your image never leaves the page.
The page loads a quantized ONNX model and runs it with
onnxruntime-web
(WebAssembly). The image is resized to 480 px, normalized, and fed to the net;
the soft-argmax head returns a normalized xyxy box, quantized to four
uint8 values — the whole detection contract. The
Mock model is matched to the bundled synthetic sample; switch to
Presence (the real-data model, reports present / absent)
and Upload a microscope photo of a threaded insert to try it.