Engineered biomechanical features — joint angles, stride length, cadence — and trained a classifier to flag form faults like overstriding and heel striking.
Diagnosed a session-leakage flaw inflating naive 80/20 accuracy to a misleading 99.7%; re-validated with leave-one-session-out CV and lifted overstride-detection F1 ~4× (0.11 → 0.43) by improving data quality.