Teach a computer to see — by snapping blocks together.
👉 Play it now: scratchml.fly.dev 👈
No code. No math. No accounts. Nothing to install.
ScratchML is a playground where kids, students, and the plain curious build a real machine-learning model the way they'd build with toy blocks:
- 🧩 Snap a recipe — stack colorful blocks like Scratch: eyes, Things to learn, a brain, a guesser
- 📸 Show examples — snap webcam pictures of each Thing… or draw them on the sketchpad if you're camera-shy
- 🧠 Train in seconds — a tiny neural network studies your examples right in the browser
- 🎉 Watch it guess — show it something new and see live confidence bars (confetti included)
The whole experience takes about two minutes from blank canvas to "whoa, it learned!" — and it's not a simulation. It's genuine transfer learning running on your own device.
- 🎨 Crayon sketchpad — 7 crayons, 3 brush sizes, eraser, undo. Train a model on your doodles; no camera required
- 🤖 A reactive robot mascot that boots up, thinks while training, and celebrates correct guesses with you
- 🔊 Toy-box sounds & music — blocks snap, captures pop, training whooshes, wins go ta-da (toggleable, of course)
- 🗑 Real dataset curation, kid-sized — every captured example is a thumbnail you can inspect and delete one by one
- 📚 It sneaks in real ML lessons: more varied examples → smarter model; the model learns whatever separates your examples — including color!
Your camera frames and drawings never leave your browser. There is no upload, no server-side ML, no account, no dataset sitting in a cloud bucket. Close the tab and it's gone.
- MobileNet v2 (self-hosted weights) as a frozen feature extractor + a small trainable head — classic transfer learning, entirely client-side via TensorFlow.js
- Smart backend selection: machines with software-only WebGL automatically fall back to the CPU backend instead of stalling on shader compilation
- The brain pre-loads while you're still reading the landing page — by the time you click Start building, it's ready
- Built with Next.js, TypeScript, Tailwind, dnd-kit, and zustand; sound and art generated with ElevenLabs and OpenAI's image models
- Tested end-to-end with a Playwright suite that actually draws shapes and verifies the model learns them
Created for Mind the Product's "Everyone Ships Now" hackathon — because the best way to celebrate shipping is to ship the thing that makes ML feel like play. #EveryoneShipsNow 🚢