Doovio watches you practise through your phone camera, counts and checks what you do, tells you exactly what to fix, and says so when it can't see well enough to judge.
Each skill has its own way of seeing and its own lessons. They all share the same honest rule: no feedback unless the camera actually saw it.
Counts squats and push-ups, checks depth, chest position and hip line from a side view, and adapts to your range once you approve it.
Coloured blocks on a printed board. Every step is verified before the next one, and mistakes are named precisely: "the blue block is on the top-middle square".
ASL fingerspelling, letter by letter, with a picture of each handshape and corrections like "bend your ring finger". A practice aid, not a replacement for Deaf teachers.
A practice talk where the camera is your audience (eye contact, hands, swaying) and a guided calm-breathing exercise checked from your chest movement.
Coming next: Piano (camera plus microphone or MIDI), Cooking, Drawing, Driving practice. They are shown as "coming soon" in the app and cannot give feedback until they really work.
A few frames per second go from the phone to the Doovio server, which sees, decides and answers within tens of milliseconds.
Claude on Amazon Bedrock is called only when the rule-based coach gets stuck. It measures the scene with OpenCV tools, decides, acts, and checks the result. It never sees an image.
Measured with the real pipeline on labelled videos and photos, including clips we never tuned on. Failures are reported too.
These are small test sets. They show how the system behaves, not statistically robust accuracy. Details, failure cases and limits are in the technical report.
A coach that watches you has to earn trust.
A five-minute walkthrough: the team, the app working, the architecture and the results.