Local LLMs, speech-to-text, vision, depth, and image utilities. Each one loads its model once and runs inference on WebGPU or WebAssembly inside your tab. No upload, no API key, no server.
Drop a video, name the words (or pick a profanity level), and get a clean cut back, with each match muted and beeped. Whisper finds the words and ffmpeg.wasm does the edit, entirely in your tab. Nothing uploaded.
Pick from Llama, Qwen, Gemma, Phi, SmolLM and more. Weights download once and cache to your browser; inference runs on WebGPU. Zero servers, zero API keys.
Index 5 public-domain novels (Austen, Shelley, Doyle, Carroll, Sun Tzu), then search them by meaning. Runs MiniLM-L6-v2 in your tab via Transformers.js.
Real-time object detection on your camera feed via YOLOs-tiny. Bounding boxes drawn on a canvas overlay; the video stream never leaves your tab.
Drop any photo, get a transparent-PNG cutout in about a second. Briaai's RMBG-1.4 segmentation model running entirely client-side via Transformers.js. No upload, no API.
Record from your mic or drop an audio file; OpenAI's Whisper transcribes in your tab with word-level timestamps. Three model sizes from 25 MB; nothing uploaded.
Pixel-wise depth maps from a 25 MB Depth Anything v2. Run it on a static image or live webcam, choose a colormap, see what your GPU sees about the world.
One million points pushed around by a compute shader on your GPU. Move the mouse to attract or repel. No model, no download. Pure WGSL.
Translate between 23 languages with Meta's NLLB-200 distilled model running fully client-side via Transformers.js. Auto-detect source via a script + stopword heuristic. No second model.
Paste text → see exactly how 6 different models (GPT-4o, GPT-4, Claude, Llama 3, Mistral, GPT-2) carve it into tokens. Color-coded chips, live char/token ratio.
Token-by-token text streaming UI. The chrome we drop in front of OpenAI, Anthropic, or local model calls.
Every demo above is a real AI application that downloads its model weights once, caches them in your browser, and then runs inference on your own hardware, usually accelerated by your GPU through WebGPU. Nothing you type, upload, record, or capture ever leaves the tab.
They're built with Transformers.js, WebLLM, and raw WGSL compute shaders, the same primitives we use when we ship production AI features for clients. Treat the sandbox as a showroom: every tile is a thing the browser can already do, with no cloud cost and no privacy compromise.
$ Want to drop in your own SPA? Build it, copy the dist into public/sandbox/<slug>/, and add an entry in src/data/sandbox.ts.