Built so a signed sentence is never lost in silence
Narvi began as a final-year Software Engineering project at Sir Syed University of Engineering & Technology, Karachi — with a simple test for success: could a deaf person sign to someone who has never learned sign language, and be understood immediately, in English or Urdu?
One goal
Sign → understood, instantly
Why this needed building
A one-way conversation
Hundreds of thousands of people in Pakistan use sign language daily — while most of the people around them can't sign a single word back.
Interpreters don't scale
Human interpreters are rare, expensive and can't be present at a shop counter, a classroom question, or a family dinner.
Phones already can
Every phone and laptop has a camera and a speaker. The missing piece was software that understands hands — fast enough to keep up with a conversation.
What a year of iteration looks like
Version one proved the idea: one hand, twenty gesture classes, a trained network hitting 86% on held-out data. It worked — but it shipped every camera frame to the server as an image, and it couldn't see a second hand at all.
Version two rebuilt the spine. The browser now sends only the hand's landmark coordinates — classification dropped to roughly 1.5 milliseconds — and the whole pipeline, from data collection to the live app, understands two hands. Gestures like namaste, clap and “we are together” finally exist in the vocabulary.
Everything runs on commodity hardware: a laptop webcam, a browser tab, and a small Flask server. No depth camera, no gloves, no GPU required.
Sir Syed University of Engineering & Technology
Final-Year Project · Department of Software Engineering · Karachi, Pakistan