How AI Golf Caddy analyzes your swing on-device (and why that matters)
When we set out to build AI Golf Caddy, we made one architectural decision early: no swing video ever leaves the phone. Here's how that works in practice.
Why on-device?
Two reasons. First, privacy: a swing video can include your face, your buddies, and the course you play at. Uploading that to a server is a lot of trust to ask for. Second, latency: cloud round-trips kill the "tap to feedback" feel. On-device runs in under a second on a modern iPhone.
The pipeline
- Capture — record swing at 60 fps using AVFoundation.
- Pose detection — Apple's Vision framework extracts a skeleton per frame.
- Swing segmentation — we identify takeaway, top, impact, and follow-through using joint velocities.
- Plane analysis — we project the club path onto the swing plane and measure deviation.
- Tempo metrics — backswing-to-downswing ratio, hip rotation timing, weight shift estimate.
- Feedback — a small CoreML model picks the most actionable cue based on the deltas.
What we don't do
- We don't upload swing video.
- We don't store swing video off-device.
- We don't sell or share your data with anyone.
Future work
The next big features are club-fit insights based on aggregated, fully anonymized launch data, and a watch companion for round-tracking that doesn't make you take out your phone.