Uploading a video of yourself mid-squat to a company you have never met is a genuinely unreasonable thing to ask of someone. So the architecture removes the question.
When you record a set, a pose-detection model loaded into your browser reads the video locally and tracks the position of your joints across every frame. What comes out is a table of numbers — hip angle, knee angle, spine angle, frame by frame. Those numbers, and nothing else, are what get scored.
The practical consequence: there is no upload to switch back on, no retention policy to trust, and no bucket of gym videos that could ever leak. The video exists on your device and then it does not.
The trade-off is honest too. Because analysis is local, it needs a reasonably modern phone and a clear side-on view. In a dark corner of a crowded gym it will sometimes tell you it cannot see you well enough — which is better than confidently scoring a set it could not read.