A fingerspelling reader that never phones home.
Most sign language tools ask you to upload video to somebody else's computer. This one does the whole job in the tab you already have open, which is both more private and, for a single hand on a webcam, entirely sufficient.
How it works, briefly
Every frame is reduced to 21 hand landmark points by MediaPipe, then normalised so that where your hand is, how big it looks and what shape your camera is stop mattering. What is left is 63 numbers describing only the shape. A small neural network of about 70 KB scores those numbers against the 24 static letters, and a letter is written down once ten consecutive frames agree. J and Z are drawn rather than held, so they are recognised from fingertip motion instead.
All of that is TensorFlow.js and WASM running on your own processor. There is no API, no queue and no account.
What it does well, and what it does not
It reads the ASL manual alphabet, one letter at a time. That is fingerspelling: names, places, words with no sign of their own. It is not a translator for ASL as a language, which uses movement, space and facial grammar that a single handshape cannot carry. Anyone telling you their webcam demo translates ASL sentences is selling something.
No headline accuracy number appears anywhere in the app on purpose. A single average says nothing about the particular hand in front of your lens, and every one of the six fist letters behaves differently across people. If it misreads you, tap the letter you meant and it adjusts for your hands, on your device only.
Who it is for
Learners drilling the alphabet, teachers who want a mirror that can actually see the front of a handshape, and anyone curious about how far on-device machine learning has come. It is free, it stays free, and there is nothing to sign up for.
Who made it
One person, built in the open about how it works rather than about who built it. Everything on the privacy page describes code that runs on your own machine, so you can verify the claims from your browser's own network panel rather than taking anyone's word for it.