Hackathon to product2026
Bridge
An accessibility companion for people whose brains, ears, eyes or hands work differently online.
- Role
- Lead engineer, team of two
- Year
- 2026
- Stack
- Next.js 16, Supabase, ONNX Runtime Web, MediaPipe, Chrome MV3, Groq
- Links
- Live product
The problem
Background music buries speech for people with ADHD, autism or hearing loss. Sign-language users can't talk to the camera on a video call and be understood. Most accessibility tools ask for a diagnosis first, then send everything to a server.
What I built
- Built at Come Build with AI (GOMYCODE × NVIDIA, 27 September 2026), then turned into a live product in English, French and Arabic.
- Sound Focus: a Chrome extension that removes background music from any tab in real time.
- Sign Bridge: sign to the camera and Bridge speaks the words.
- An AI guide, plain-language courses, Simplify for dense pages, and Page check.
How it earns trust
On device by design
Sound Focus runs Silero VAD, YAMNet and a GTCRN speech enhancer in ONNX Runtime Web inside the extension, in about 150 ms. Tab audio never leaves the computer.
An eval for the audio router
A pipeline replays labelled audio fixtures and scores how much speech survives and how much music leaks, against a frozen copy of the old router. Regressions show up as numbers, not impressions.
Row-level security, tested
19 end-to-end checks create temporary creator, buyer and guest users, try to cross the boundaries, then clean up after themselves.
Validated model output
The guide works with any OpenAI-compatible provider with automatic fallback, and its output is validated before anyone sees it.
By the numbers
- music removal, fully on device
- ~150 msmusic removal, fully on device
- automated RLS checks
- 19automated RLS checks
- languages, with right-to-left
- 3languages, with right-to-left
- from hackathon to live product
- 3 daysfrom hackathon to live product
How it fits together
- Tab audiois there speech?Silero VAD
- Silero VADis there music?YAMNet
- YAMNetrouteGTCRN enhancer
- GTCRN enhancerclean speechSpeakers
What I'd do next
Publish the audio eval fixtures and scores, so anyone can check the router's claims for themselves.