Ship the feature extraction for both runtimes, not just Python

Document both runtimes, and how to decode audio the way the model was trained

The detector, its normalisation, and an honest card

initial commit