This project tracks body landmarks with MediaPipe and can generate spoken coaching by interpreting external ML JSON through Gemini, then sending text to ElevenLabs.

python3 body_tracker.py

Controls:

• q: quit
• v: toggle continuous coaching (repeats until desired position is reached)

Use a .env file in the project root with plain KEY=VALUE lines:

GEMINI_API_KEY=your_gemini_key
GEMINI_MODEL=gemini-1.5-flash

ELEVENLABS_API_KEY=your_elevenlabs_key
ELEVENLABS_VOICE_ID=your_voice_id
ELEVENLABS_MODEL=eleven_multilingual_v2

# Required for coaching: path to ML output JSON consumed by Gemini prompt
ML_JUDGEMENT_PATH=/absolute/path/to/ml_judgement.json
COACHING_INTERVAL_SEC=3.0

If keys are missing, the app still runs and falls back to local text feedback (no voice).

When you start the app, it prints:

• [Config] Gemini configured: yes/no
• [Config] ElevenLabs configured: yes/no

If voice fails on v, terminal logs now include the exact ElevenLabs error.

Any schema is allowed. Gemini is prompted to infer a human coaching line from your model output.

{
  "joint": "L.LEG",
  "direction": "LEFT",
  "magnitude": 1.0
}

Possible response: Move your left leg slightly inward.

Coaching continues until ML JSON indicates the desired position was reached. Supported flags in your JSON:

• desired_position: true
• target_reached: true
• position_correct: true
• is_correct: true
• status: "correct" (also accepts "aligned", "good", "ready", "ok")

Install Python packages used by the app:

pip install mediapipe opencv-python requests