Not every user has fast internet, and not every feature should send data to a server. In this station, participants add an on-device AI feature to the starter app using Transformers.js. The model downloads once, runs in the browser on WebGPU, and falls back to WASM on weaker devices. You’ll learn how to pick a model size for the device in front of you, how caching works so the download only happens once, and where small models fall short. The pattern comes from an offline recipe generator I built that runs entirely in the browser with no server round-trip.

What participants build: an offline AI feature (such as summarising or classifying text) with a model loader, a progress indicator and a WebGPU-to-WASM fallback.

What they take home: a working in-browser inference setup, plus a rule of thumb for choosing between small and larger models per device.

Prerequisites: basic React or JavaScript. A laptop with a recent Chrome or Edge is recommended.

Notes for organisers: the first model download is the slow part. I can point to a smaller model for low-bandwidth venues. Please confirm that the venue Wi-Fi can handle it.