Sketch Lab
The training side of this site's ML Lab and Attention Lab. A shared conv trunk powers two models: a digit recognizer (99.4% MNIST test accuracy, reads multi-digit numbers via connected-component segmentation) and a 75-class doodle classifier trained on over a million Quick, Draw! sketches; a separate X-ray export exposes every intermediate tensor so the site can render the CNN thinking live. Alongside them, a 620k-parameter character-level GPT trained from scratch on Shakespeare, with hand-written attention so the softmax(QK^T) tensor exports as a first-class ONNX output for the Attention Lab. Normalization and softmax are baked into the graphs, the browser preprocessing reproduces MNIST's own crop/scale/center construction, and parity gates cover every export. Inference runs client-side via onnxruntime-web (WebAssembly): nothing you draw or type leaves the page.