Client projectAILive
Wellnix
AI health product with two working modules: MuscleAI reviews lifting form from workout video, and NutriAI reads a nutrition label and scores it against your goals.
Result
Form review and label scanning in one product
The problem
Working out at home means nobody tells you your form is wrong until something hurts.
What I built
Built a computer-vision pipeline that reviews posture and movement from ordinary phone footage, and a label reader that extracts ingredients and nutrition panels and scores them against the user's goals and dietary pattern. Both feed one profile.
The outcome
Form feedback without a trainer in the room, and a clear consumability score with its reasoning for any packaged food, adapted to the person reading it.
How it works
- 01
The user records a movement on an ordinary phone camera, with no wearable and no depth sensor.
- 02
Pose estimation extracts joint positions per frame and the movement is scored against the shape the exercise should have.
- 03
Feedback comes back as the specific correction, not a score, because a number does not tell anyone what to change.
- 04
A photo of a nutrition label is read for ingredients and panel values and returned as a consumability score with the reasoning behind it, shaped by the user's goals and health context.
The AI layer
Pose estimation on video frames, then rule-based scoring of the extracted skeleton against per-exercise form criteria. The scoring is deliberately not a model: form rules are known, explainable, and easier to correct when a coach disagrees with them. On the nutrition side, label extraction feeds a model that explains the score in terms of the user's own profile.
The engineering layer
Client-side capture with server-side analysis, frame sampling to keep it affordable, and a data model that keeps movement and nutrition in one profile rather than two apps.
Key technical decisions
Deterministic scoring on top of a learned model.
Pose estimation is the part only a model can do. Judging whether a knee tracked past the toe is arithmetic, and doing it in code makes every judgement explainable to the user.