A small collection of interactive, in-browser explainers I built to help make core ideas in machine learning and AI more intuitive. Each one is a self-contained page you can click through at your own pace — no installation, no code, just open and explore.

Available explainers

  • How a neuron learns — Watch a single neuron learn to detect “is the number big?”, step by step, as it adjusts its weights through training.

  • From text to embeddings — The full pipeline that turns a sentence step by step into the numbers an AI model can understand, ending in dense embeddings.

  • Perceptron visualiser — A live Streamlit app (running entirely in your browser via stlite) where you set the inputs, weights, bias and activation function of a single neuron and watch how its output changes. Note: the first load takes a few seconds while the Python environment downloads.

  • Transformer architecture explorer — A live Streamlit app (also via stlite) that walks through the components of the Transformer architecture — embeddings, positional encoding, multi-head attention, feed-forward layers and more — with a short quiz to test your understanding. The first load takes a few seconds.

  • The Lego way of building CNNs — Snap together a convolutional neural network for MNIST like Lego bricks: stack and reorder convolution, max-pooling, flatten, linear and softmax layers, setting every parameter by hand, then pick your learning rate, epochs, loss function and optimizer. Each run is a simulated training curve; you get five iterations, a leaderboard ranking them by test accuracy, and the PyTorch code for the model you built.

  • Going back in time: run GPT-2 like it’s 2019 — Type a prompt and watch GPT-2 continue it, token by token, with sliders for temperature, top-k, repetition penalty and length. The 124M-parameter model itself runs on your own machine: a quantised ONNX export executed by Transformers.js and ONNX Runtime Web (WebAssembly), in a Web Worker so the page stays responsive. No server, no API key — your prompt never leaves your laptop. The first run downloads about 128 MB of weights, which the browser then caches.

  • Convolutions with MNIST — Work through a convolution one step at a time on a real handwritten digit: pick an MNIST image, turn it into the 28 × 28 matrix of pixel values the network actually sees, choose a 3 × 3 or 5 × 5 kernel (vertical or horizontal line detection, blur, sharpen, outline — or type your own weights), and see the resulting feature map. Hovering any output pixel highlights the patch it came from and spells out the multiply-and-add behind it. The digits are real MNIST test images exported with torchvision, so the page needs no install and no server.