The AI that watches your form
This prototype demonstrates the intended loop: Harry Spotter counts reps, maps rubric findings to a specific rep, and passes validated metrics to the deterministic training engine.
Harry Spotter is analyzing your technique…
Two planned tiers
Local feedback first. Cloud analysis by request.
These cards describe the target architecture, not features available in #13. Device and cloud validation land in later integration changes.
Tier 0 — On-device target
- Target: skeleton overlay at camera framerate
- Target: rep counting + range-of-motion tracking
- Target: basic mid-set fault flags
- Designed for local inference without per-set cloud cost
Tier 1 — Cloud analysis target
- Target: frame-by-frame cloud analysis of an explicitly chosen clip
- Target: versioned rubric per supported exercise
- Target: keyframe report with findings pinned to a rep
- Target: generated explanation constrained to engine findings
Severity that means something
Tips, warnings, and the one red we never waste
The planned detector reads versioned rubric thresholds rather than inventing findings. Three severities map to three deliberately different responses.
“Depth above parallel on reps 5–6”
Technique polish. Delivered kindly, with a drill (goblet pause squats) queued for your next warm-up. Tone dial applies — the wizard may editorialize.
“Knees caving inward on reps 3–5”
Worth fixing before adding load. The engine may hold your progression on that lift until the pattern improves — and it will tell you that's why.
Voldelift appears
Rare by design — that's what makes it effective. The visual can be theatrical; the words are always plain language:
Your lower back is rounding under load. End the set and reduce the weight. If you feel pain, stop training and consult a professional.
The report
A form score you can argue with
Not a black box: the score decomposes into findings, each finding pins to reps and keyframes, and every threshold lives in a versioned rubric file.
Rep-quality gate: a set of eight where six reps cut depth isn't a set of eight. The progression engine only counts what the camera can defend.
Form Ghost: your best-form rep replays as a ghost skeleton over a later set. This interaction remains a planned vision feature.