LVL-UP SPORTS AI
- 0.94
- F1 ball detection
- 60k+
- labelled frames
- 85–90%
- stroke accuracy
- ₹60
- GPU inference / hour
All details: LVL-UP SPORTS AI
Additional metric
- ₹60 — GPU inference / hour
- Managed a four-member team and delivered the racquet-sport analytics product end-to-end, from prototype to on-court use.
- Design and deliver an end-to-end computer-vision pipeline for racquet-sport match analysis, spanning ball and player tracking, court calibration, pose-based stroke classification, and rally segmentation.
- Established live-court data acquisition and an annotation workflow using on-court camera installations and multi-camera stereo setups; curated a 60,000+ frame labelled dataset for ball, bounce, and stroke events.
- Trained detection and classification models achieving 0.94 F1 for ball detection on held-out match footage; pose-based stroke detection sustains 85–90% accuracy on live match play.
- Architected a modular analytics engine combining six perception streams (court, players, pose, ball, racket, swing) into shot- and rally-level events with confidence scores and physics-based validation.
- Optimized GPU inference economics to approximately ₹60 (~$0.70) per hour of footage.
FID · Harsha wired me with YOLO, TrackNet and event-based classification to track court, players, pose, ball, racket and swing. I call every bounce. Still can't serve.
