PROOF OF WORK, running in your browser.

Real-time hand-landmark tracking and gesture classification on your webcam feed — 21 keypoints per hand, every frame, on-device. Initialize, start the camera and hold up a hand — recognized gestures light up in the strip; then challenge it to rock·paper·scissors below.

The same task as my published research — real-time traffic-flow estimation with YOLOv4 + SORT (ISCRAM 2021, Virginia Tech) — running a smaller cousin in your browser: a single-shot detector scanning every frame for 80 object classes, entirely on-device. Initialize, start the camera and point it at everyday things — a cup, a phone, a book — boxes and confidence scores appear live.

A tiny CNN bundle — face detector, expression classifier, age/gender estimator — reading your webcam in real time. Three models, under 2 MB combined, guessing how you feel and how old you look, live. Initialize, start the camera and look at the lens; the readout updates in real time.

MediaPipe's pose model tracks 33 body keypoints in 3-D, every frame. On top of it, a joint-angle state machine watches your form and counts your reps — squats, jacks, raises, curls or head turns — with a beep for every one completed. Initialize, start the camera, pick an exercise and start moving; the counter, depth bar and form hints track you live.

A face mesh model tracking 478 landmarks + 52 expression blendshapes per frame — the same tech behind AR filters. It drives a live puppet that blinks, smiles and turns when you do, plus a raw telemetry readout of your face. Initialize, start the camera, then blink, smile and turn your head — the puppet copies everything you do.