What we build
Core design
Three decisions define this product.
Privacy by architecture
Camera frames never leave your phone. Pose detection runs on-device; only numeric joint angles (7 angles, ~50 floats per frame) are sent to the LLM for voice coaching.
24-form methodology
Built around the traditional 24-form Tai Chi curriculum, with real-time scoring on joint positions and personalized weekly coaching plans.
Master Wang lineage
30 years of teaching experience encoded into the form standard. The AI coach voice mirrors the pace and tone of a patient human instructor.
How it works
How it works
From detection to voice coaching — five steps, all under privacy constraints.
On-device pose detection
MediaPipe Pose Landmarker (9MB model) runs locally on your phone's GPU. No frames uploaded.
Real-time scoring
7 joint angles compared against the 24-form reference template. Score updates 30 times per second.
AI voice coaching
Numeric scores only are sent to the LLM (via Cloudflare Worker proxy). The LLM returns personalized corrections in natural language.
Text-to-speech
Coaching feedback is rendered in a natural-sounding Mandarin voice (speech-2.8-hd, "Humorous_Elder").
Your data stays yours
Practice history is stored locally. Clear it any time in app settings.
Contact
Get in touch
Pick the mailbox that matches your question — all answered personally.