Locate the body
On-device pose estimation turns visible joints into time-stamped landmarks. The source image is not sent to an AI model for scoring.
Method & limits
Feedback built from movement, not a black-box verdict.
Patchworks separates rep detection from form scoring. It reports what the camera could measure, leaves unavailable categories ungraded, and keeps the limits of a single-camera view visible.
Set up an analysisFollow the movement
Exercise-specific state machines look for a complete down-and-up or extension-and-return cycle while filtering incidental motion.
Measure the rep
Each completed rep is evaluated against movement-specific biomechanics features. Unavailable measurements do not lower the score.
What changes the result
The camera is part of the measurement.
A clear view of the requested joints matters more than a cinematic frame. Squats work best at a three-quarter angle with both legs visible. Pushups work from a side view and can use the clearer side when the far arm is hidden.
- Frame the complete movement. Keep the requested body regions visible from start to finish.
- Treat missing data honestly. Occluded or low-confidence measurements are marked unavailable instead of guessed.
- Read trends, not diagnoses. Scores help compare observable form across controlled sets; they do not assess pain, injury risk, or readiness.
Privacy model
Analysis stays local. Playback is opt-in and temporary.
Landmark analysis and scoring run in your browser. If you choose rep playback, the recording is uploaded to private temporary storage for the active session and removed when you end it, with automatic cleanup as a fallback.