Sign-See

Sign-See started as a two-week challenge set by a Dutch disability center that works with sign language. The brief was open: build something that could help people learn NGT. Ana and I built the first working backbone, and in a later deployment block the team rebuilt it into a real application. I led the frontend. At the end of year two, the four of us chose to keep going and turn Sign-See into a startup.

Today I am a co-founder and product lead, responsible for how Sign-See looks, feels, and behaves for the people who use it. The gap we are addressing is real: most deaf children in the Netherlands have hearing parents who need to learn NGT quickly, while current options are largely scheduled courses and passive video.

At the heart of the app, a camera composable captures webcam frames at around 10 fps, streams them to the backend over WebSocket, and draws a 21-point hand skeleton over every response. Backpressure keeps no more than one frame in flight, while automatic reconnection helps the experience remain smooth when the connection does not.

• Four connected product modes: Learn, with hints and hold-to-validate practice; Collect, for contributing labeled examples; Home, which ties the experience together; and Game, which gamifies practice across the app so learning feels engaging, progress is visible, and people want to return and continue.
• A 26-test frontend suite and a file-by-file WCAG 2.2 AA accessibility retrofit.
• Production work including an Azure database migration, CI fixes, and a collaborative PR review cycle.

The team’s recognition pipeline uses MediaPipe hand landmarks and a fine-tuned EfficientNet-B0, with a landmark-based MLP supporting low-confidence frames and temporal smoothing preventing flicker. A FastAPI service exposes HTTP and WebSocket endpoints with JWT authentication and PostgreSQL storage. Around it sits an MLOps stack spanning Azure ML, MLflow, Airflow, a versioned model registry, and CI/CD.