Sprint form analysis without a server
Can a phone analyse sprint mechanics well enough to coach from, without uploading the video anywhere?
Started · Updated
The obvious architecture for video analysis is to upload the clip and process it in the cloud. It is also the architecture with the worst properties for this particular product: an athlete records on a track with poor signal, the clips are large, the processing cost scales linearly with use, and the footage is of a person’s body.
SPRIVO runs the pose detection natively on the device instead, through a MediaPipe integration rather than a hosted inference service. The sprint never leaves the phone to be scored.
The open question is not whether it runs — it does — but where the accuracy ceiling sits for the specific mechanical faults worth coaching, and how far the analysis can go before a cloud model would genuinely be better. That is what keeps this in the Lab while the product itself is in closed testing.
What we learned
Findings so far.
- 01
On-device analysis removes a whole category of problems at once: no upload wait on a bad connection, no per-clip processing bill, and no server holding video of somebody’s body.
- 02
It moves the cost from infrastructure to integration. The native layer is the substantial engineering, and it is platform-specific in a way a hosted API would not have been.
About SPRIVO
This experiment came out of building SPRIVO.