Making movement legible

Most people can record themselves moving, but a video alone does not make it easy to see what changed from one repetition to the next. I built threedpt as a browser-based instrument for turning that footage into something a person can inspect.

It captures movement through an ordinary webcam, tracks 33 pose landmarks with MediaPipe, and reconstructs the motion as a scrub-able 3D body. The analysis surfaces range of motion, left-right asymmetry, tempo, rep count, and estimated joint load. A second view lets someone mark where they felt pain and inspect the anatomy around that joint at the relevant moment.

Trust as a product constraint

This is health-adjacent software, so the analysis engine is deterministic rather than generative. Every observation shown in the interface is meant to trace back to a measured value instead of an LLM interpretation.

That does not make a single-camera estimate clinical truth. Depth, occlusion, camera position, and incomplete tracking all place limits on what the system can know. The project is an observation and learning tool, not a diagnostic device or a replacement for a clinician.

The system

The application runs on-device in Next.js and React. MediaPipe handles pose landmarking, a TypeScript analysis pipeline smooths and measures the motion, and Three.js powers the interactive body and review environment. Recordings and history remain local to the browser.