Climbing training with technique analysis through computer vision, working offline.

An installable climbing training app: routines, goals by route color, nutrition tracking and a module that analyzes technique from a video of your own climb. The pose estimation model runs entirely on the phone, over 33 body points, and returns technique scores without the video ever leaving the device.
The core decision was for everything to work offline, and that forced solving three things that turned out to be the most transferable lessons of the set. Video is captured with the browser's file picker instead of asking for direct camera access, because the first works on a local network without a certificate and the second requires HTTPS. The model weighs almost 6 MB, so it is left out of the initial install and downloaded the first time the module is used: installing the app stays instant. And all the pose-to-score logic lives isolated, with no interface around it, so it can be reasoned about and fixed on its own.
There is one product decision worth telling: the streak is weekly, not daily. A daily streak in a training app pushes you to train injured so you do not lose it.



