Coaching that happens in the same breath as the movement.
The fitness unicorn's flagship feature was an AI coach that critiqued form in real time. Beautifully designed, universally beloved in demos, and utterly unusable in a real workout — the round-trip to a cloud GPU took 900 milliseconds, which is exactly long enough to break the flow of a repetition.
We rewrote the inference core in Rust and compiled it into a shared library that both iOS (via CoreML) and Android (via TFLite) could consume through FFI. The client was rebuilt in Flutter with a thin platform layer, so 92% of the app is shared code and 100% of the coaching path runs on-device.
The trailhead problem — no signal, real workout — became the design pillar. Model updates are pushed through a signed OTA channel; telemetry batches and syncs opportunistically. If the network never comes back, the coach still works.
Median feedback latency is now 58 milliseconds. Thirty-day retention is up 27%. The app was featured on the App Store in fourteen countries in the quarter after launch.
“It stopped feeling like a product and started feeling like a spotter.”
Kepler's on-device pipeline is now being extended into a form-correction studio for physiotherapy, in a joint pilot with two European clinics.