A fitness feature that waits on the cloud for every rep feels broken, and sending a continuous camera stream off the device is a privacy problem members did not ask for. The useful version runs on the phone, degrades clearly when it is unsure, and keeps the coaching tied to what the camera actually saw.
What makes AI work different here
- Latency is the product. Form feedback that arrives late is wrong even when the classification is right. On-device inference is the default, not an optimization you add later.
- The camera feed is sensitive. Workout video is personal data. Keeping inference on the device avoids a stream of frames you then have to secure, retain, and explain.
- The model has to admit uncertainty. A confident correction on a bad pose estimate teaches the wrong movement. The app should fall back when the pose is unclear.
- Phones vary. The same model has to behave on last year's mid-range phone, not only on the device used in the demo.
Where we typically start
We start from the moment in the workout the model has to answer, the latency that moment can tolerate, and what must stay on the device. Cloud fallback, if any, is a separate path with its own privacy review.
Services we typically provide
Case studies from this industry
Frequently asked questions
Does the video have to go to the cloud?
Not for the core check. The published fitness work runs inference on the device so the camera stream is not the product's data path.
What if the phone cannot keep up?
The feature should say it cannot see the pose, rather than invent a correction. We size the model for the devices you support, and we measure that before launch.
Can this coach nutrition or medical advice?
No. Form and workout guidance is the scope. Health claims and medical advice are a different product with different review, and we do not fold them into a pose model.