Trefoil Detector

An end-to-end nuclear-trefoil detector — trained on synthetic data, exported to ONNX, accelerated in the browser via WebGPU/WebGL. 37/37 real images detected, 0 false positives.

Portfolio piece for an IAEA Department of Safeguards-flavoured AI/ML contracting role.

⚛ Demo project Live in-browser detection, real images at random 3D angles 📝 Journal & writeup How I built it — the polished writeup and the raw engineering log

Stack

Training: PyTorch + Ultralytics YOLOv8n on Apple MPS
Data: 11k synthetic images across four evaluation-driven generation rounds (programmatic compositing onto COCO backgrounds); 37 hand-curated real images held out for validation
Export: ONNX (opset 13), ~12 MB at 416×416
Runtime: ONNX Runtime Web — WebGPU primary, WebGL fallback, WASM final fallback
Demo: Three.js fly-by — detect once on the clean frame, box tracked by camera projection; live confidence gauge + 3D mechanical odometer