Runs entirely in your browser — no video leaves this machine. Hand tracking by MediaPipe Tasks Vision.
Recognition uses a k-nearest-neighbour classifier over normalised two-hand landmarks; swap in a dataset-trained model at classify().