Reading sign language from a webcam, 30 frames a second

A webcam sees a hand; the system has about 33 milliseconds to say what it means. MediaPipe extracts 21 landmarks per hand, then a feature layer turns those raw coordinates into 48 geometric descriptors — fingertip-to-wrist distances, inter-finger angles, hand orientation — so the classifier learns gesture shape rather than where the hand happened to sit in frame.

A random forest over those features hits 92-96% accuracy at roughly 1-2ms per inference. The harder problem was stability: frame-by-frame prediction flickers badly during fast movement, so a 5-frame majority vote smooths the output and suppresses the confident-but-wrong predictions that appear mid-gesture. Built end to end from scratch — no pretrained gesture model.