AI · computer vision
On-device AI & recognition
Camera and voice features that run on the phone for the common path, with cloud fallback only when confidence or policy requires it.
KotlinML KitCameraXOn-device ML
On-device
first, cloud when needed
The problem
Cloud-only inference is slow and expensive on mid-range hardware and patchy networks; capture quality often matters more than the model.
What we did
- 01Run the common path on-device (ML Kit / voice models); reserve cloud for low-confidence cases
- 02Tune capture UX — framing, retake prompts — before chasing model size
- 03Measure accuracy, latency, and cost per request as one budget before rollout
Outcome
- Recognition in seconds offline for the common case
- Lower per-request inference cost vs cloud-only
- Ship-ready evaluation harness for accuracy and spend