Case 01 · Computer vision · Full-stack
SakhiSign
An AI platform that tells you whether you signed it right — and which part was off.
The problem
Checking whether a sign-language gesture was performed correctly usually means training a classifier on a large labelled dataset — one that has to grow every time a new sign is added. And a plain pass/fail doesn't tell a learner what to fix.
What I built
- A zero-training-data evaluation engine: each sign needs just one reference recording.
- Real-time tracking with MediaPipe Hands — 21 landmarks per frame turned into an 18-D feature vector.
- Dynamic Time Warping to compare a learner's attempt against the reference, even at a different speed.
- Mirror-invariant scoring and targeted, per-component feedback on each attempt.
Key decisions
- Compare, don't classify. Matching against a reference recording means adding a sign is a recording, not a retraining run.
- Split the services. Auth (JWT) and data (MongoDB) sit behind Node/Express; inference runs in a separate FastAPI service.
21landmarks tracked per frame
18-Dfeature vector per frame
0training samples needed
3tiers, deployed live