Handful - ASL Learning App

Inspiration

I wanted to learn ASL, but the good apps cost money and the free route was a pile of YouTube videos with no order, no way to check what I'd remembered, and nothing to tell me if I was signing correctly. Handful puts it all in one free place.

What it does

Handful teaches ASL in short lessons: 72 signs across 24 lessons, with real signer videos from ASL Signbank, handshape drawings and memory tips. You practise with quizzes, matching games, spaced review and a mistakes list, with streaks, XP and a daily goal. A camera check gives practice feedback on all 26 fingerspelling letters and 3 words (EIGHT, WHY, YES). Spell it shows the handshape for any name or word you type. There's no account, no backend and no database. Progress is saved in your browser.

How AI is used

There is no chatbot. The AI is computer vision running entirely in your browser, so your camera video never leaves your device.

• Google MediaPipe Hand Landmarker finds 21 points on your hand in every frame.
• A small logistic-regression model, trained with Claude Code on about 2,100 public hand recordings, tells apart the fist letters A, E, M, N, S and T by where the thumb sits (about 92% on sessions it had never seen).
• Matching logic compares your hand shape and movement against each letter's requirements, and against measurements taken from Signbank videos for words.

Safety and limitations

The biggest risk was telling a learner they're right when they're wrong. I waved a hand at the camera and it accepted it as BLUE. I then measured the false matches, gave each word its own strict limit, required moving signs to actually move, and switched the camera off for words it couldn't check reliably. A thumbs up was also accepted as YES, which I fixed with a thumb check. The app tells learners plainly that it's practice feedback, not a grade. Limits: the word check covers 3 words and has only tested on me. The letter check is weakest on a few letters, and it follows one hand only. The sign descriptions were drafted by a non-signer and haven't been reviewed by a fluent ASL signer yet, and the app says so.

Built with

Next.js, React, TypeScript, Tailwind, MediaPipe Tasks Vision, ASL Signbank videos (CC BY-NC-SA 4.0, so Handful stays free and non-commercial), and Claude Code as the AI coding assistant.

What's next

Test with many more people, hands and lighting, get a fluent ASL signer to review every sign description, and add more words once they hold up.

Built With

• claude-code
• computer-vision
• machine-learning
• mediapipe
• nextjs
• react
• tailwindcss
• typescript
• vercel