Biology is complex but kinematics does not have to be. A handful of angles (wrist, elbow, shoulder; toe, heel, ankle) over time capture complete human movement. That data is manageable enough to read and process.
An early experiment in this area was to record pull-ups with a webcam and run the footage through a pose estimation (mo-cap) model to produce 3D coordinates for each joint. I fed the data to a frontier LLM and the raw matrix was far too unwieldy, but processed mathematically (isolating the joints at key muscle groups, taking min/max values…) it gave exactly the information needed for action quality analysis. Training based on the LLM’s feedback improved the pull-ups. The less powerful the model, the more watertight the math must be since the model will be less capable of drawing conclusions from complexity.
[Jabhook](../) is the implementation of this, as a negative
feedback loop for boxing: Call out combination → Observe combination
→ Analyze/improve combination → Repeat. How the Jabhook
biomechanics algorithm works (metaphorically): tracing paper with perfect
form on it, held up to the user. The algorithm finds the gap between the
outlines and explains exactly how to close it. Most education runs on
negative feedback (teachers are amazing though limited by working hours)
but Jabhook is so fast, specific and actionable that it sculpts form.