CT Learn · AI Coding · Level 4

Your hand is the controller.

A neural network finds 21 landmarks on your hand, sixty times a second — live in this tab. Draw in the air with a fingertip. Pinch to switch color. Make a fist to wipe the canvas.

All AI runs in your browser — nothing is uploaded.

  • 21 Hand Landmarks
  • Gesture Recognition
  • 60 fps
  • On-device
HAND_LANDMARKER · 21 KEYPOINTS
p8 · index_tip  x 0.512  y 0.148  z -0.041  conf 0.98

hand_landmarker.task · float16 · ≈12 MB · runs on your GPU

Pick a mission

Eight ways to bend reality with one hand

Same neural network, same 21 landmarks — eight different universes built on top.

How it works

Three steps, sixty times a second

01

Camera

Your webcam streams frames straight into the page. Nothing is recorded and nothing leaves this tab.

02

Neural network

A 12 MB hand-landmark model runs right here — WebAssembly plus your GPU, about sixty predictions every second.

03

Landmarks

Each prediction is 21 (x, y, z) points per hand. Every gesture you'll see is just geometry on those points.

Demo built for CT Learn — AI Coding · Level 4 · Project 1 · Model: MediaPipe Hand Landmarker

no accounts · no api keys · no cloud — the model lives in your tab