MediaPipe Pose Landmarker
A pretrained 33-point body pose model (Heavy or Full variant) runs the movement tracking. It is the authoritative source of landmarks. Model files are served from this site and cached locally.
AI Stack
Real machine learning. Running on your laptop.
Nothing here is a gimmick. Pretrained pose models do the seeing; a transparent engine does the counting; and the two are kept honest by an uncertainty layer that would rather say nothing than guess.
MoveNet Thunder
An independent second pose model cross-checks a handful of shared lower-body joints. If it fails to load or disagrees, counting pauses or falls back to the primary model — it can never break the core experience.
Deterministic movement engine
Repetitions and holds are decided by transparent, rule-based state machines with calibrated thresholds — not by a neural net guessing a number. Every count has an auditable reason.
Uncertainty engine
A weighted confidence score combines landmark visibility, jitter, freshness, camera view, and model agreement. Below threshold, the app abstains instead of inventing a rep.
On-device Prompt API (optional)
When your browser exposes a built-in language model, an optional plain-language recap is generated locally. When it is absent, a deterministic recap is used instead. No text is sent to a server.
The pipeline
Every frame takes the same eight steps.
Camera pixels enter a worker, become landmarks, and leave as movement — the pixels themselves never come back out.
- 01CameraLocal frames only
- 02MediaPipe Heavy/FullPrimary pose AI
- 0333 landmarksBody coordinates
- 04MoveNet verifierExperimental
- 05Uncertainty fusionConfidence-aware
- 06Temporal engineSession template
- 07Rep / hold stateDeterministic count
- 08Local feedbackOn your device
Boundaries
What the AI does not do
- It does not train or fine-tune on you. The models ship pretrained and never learn from your sessions.
- It does not judge whether your movement is medically appropriate.
- It does not upload frames, landmarks, or recaps.
- It does not decide a number it cannot justify — low confidence means it abstains.
Curious about the second model in action? Open the AI Lab to watch the skeleton and the verifier side by side, or read the research behind the routines.