Velo
Tennis coaching feedback from a single phone video.
- Status
- Collaborative hackathon prototype
- Year
- 2026
- Role
- Pose estimation, with Craig Mutugi
- Area
- Machine learning
In plain English
Good coaching usually needs a coach watching in person. Velo lets a player film one clip on a phone and get feedback on how their body moves through each stroke. It also keeps a verifiable record of their coaching history.
How it works
- A player records one tennis clip on a phone.
- Pose estimation (software that finds body joints in video) tracks the player's movement frame by frame.
- AI agents turn that movement into biomechanical feedback (notes on how the body moves).
- The feedback is recorded on-chain (on a blockchain ledger) so the coaching history can be verified.
What I did
Craig Mutugi and I built it for the Somnia Agentathon. I worked on pose estimation with YOLO, and Craig built the Somnia agents and the MediaPipe pipeline.
What I learned
I tried several fine-tuning approaches, and the stock model handled awkward camera angles, blocked views and real match footage best, so we shipped it. Knowing what to leave alone was the most useful result.