Tracking human movement, joint by joint

We ran keypoint detection on a video of someone running — MAKRR tracking every major joint, frame by frame, no manual labeling.

The model doesn't know what it's looking at until you tell it. Running, walking, reaching, bending — the joints behave the same regardless of context. That's what makes it useful well outside where the original footage came from.

The same setup could:

• Flag unsafe lifting posture in a warehouse before it becomes a back injury

• Track gait changes in an assisted-living facility for early signs of fall risk

• Catch a construction worker reaching without support

• Measure physical therapy recovery week over week, without a clinician manually angling a goniometer at each visit

Every time we demo this to someone new, we come away with a couple more use cases.