Case study · Team project

AirBlocks

A block puzzle you play with your bare hands in the air — webcam hand tracking replaces the mouse entirely.

MediaPipeOpenCVFlaskJavaScript
AirBlocks — Loop — hand tracking to grid placement

01 — Problem

Gesture control only feels good if the grab is unambiguous and the latency is invisible. The first build ran the webcam and the MediaPipe loop on the server through OpenCV, which works on your own machine and nowhere else — a cloud host has no camera attached, so the game could not be deployed or shared at all.

02 — Approach

MediaPipe Hand Landmarker tracks 21 landmarks per hand; a closed fist or a pinch grabs a block and an open hand drops it onto a 10x10 grid, with an adjustable region-of-interest so players can frame the exact space their hand moves in. The deployment problem was solved by shipping a second edition: the same game rebuilt fully client-side with MediaPipe Tasks for Web, so camera and inference both run in the visitor's browser and the whole thing hosts as static files.

03 — My role

Team project for a Computer Vision course. I built the web interface and owned deployment — including the client-side rebuild that made a free public demo possible.

Team project

Results

Two working editions from one codebase — a Flask build for local experiments and metrics capture, and a browser-only build that deploys free to static hosting and is the one people can actually try. The Flask edition logs average FPS and grab/drop accuracy to CSV for evaluation. Two grab gestures and a camera selector made it usable across different machines and lighting during testing.

21

Hand landmarks

10x10

Grid

2

Deployable editions

Demo recording

Screen recording — grab, move and drop with an open hand

Next projectCekPKWT