in-browser-ml

PyTorch's API that trains in a browser tab on WebGPU — autograd, nn, optim, capture/replay, int8, ONNX out; held to real torch's values on 4,764 golden cases; measured on one page against TF.js, jax-js, Burn and ONNX Runtime Web (4–12× faster training, ahead of ORT at inference)

Browser palm-crease segmentation: a 5.5M-param ONNX UNet that traces heart/head/life lines client-side. Live demo + full training pipeline.

Multi-label emotion classification on GoEmotions: a 3.8 MB linear model at 96% of BERT, served from FastAPI and in-browser JS

Do you need a transformer to forecast? Five tiny models (none over 80k params) vs published transformer results on the ETTh1 long-horizon benchmark — with an in-browser demo.

Upload any CSV, get a trained + evaluated + explained model. Real scikit-learn running in your browser via Pyodide — no backend.

AI-powered tweet sentiment classification dashboard with in-browser ML inference, built with Logistic Regression, TF-IDF, and Chart.js on GitHub Pages. Zero server required.