Build it yourself.
12 hands-on projects, from a tiny language model to fine-tuning and shipping your own API. Each is broken into steps with working code and a check to tell you it's right, and every code step has been run and tested.
A tiny language model
Generate new names by predicting one letter at a time, the same idea as an LLM predicting tokens.
Uses: Tokens and embeddings, Transformers4 stepsBuild a tokenizer
Teach a program to split text into tokens the way GPT-style models do, by learning merges from a book.
Uses: Tokens and embeddings, ML from scratch5 stepsBuild a spam filter
Train logistic regression from scratch on real text messages, then decide how strict the filter should be.
Uses: Classical machine learning, Python and maths toolkit5 stepsA search engine for your notes
Build BM25, the keyword search behind most search boxes, and point it at your own notes.
Uses: Search and retrieval5 stepsAn A/B test analyser
Write the maths behind an experiment dashboard: significance, confidence intervals, sample size and a bug check.
Uses: Statistics and experiments4 stepsA neural network in NumPy
Write the forward pass, backpropagation and training loop yourself, and prove your gradients are right.
Uses: Neural networks, Python and maths toolkit5 stepsAn evaluation harness
Build the tool every AI team needs: run a fixed test set, see what passed, and know whether a change really helped.
Uses: Evaluation, Prompting and context5 stepsChat with your documents
Build retrieval-augmented generation end to end: chunking, retrieval, a grounded prompt, citations and an evaluation.
Uses: Search and retrieval, Prompting and context, Evaluation6 stepsA tool-using agent
Build the agent loop yourself: tools, a step limit, a trace of every action, and tests that don't need an API.
Uses: Agents, Prompting and context, Production AI and MLOps5 stepsRecognise handwriting with PyTorch
Train a small convolutional network that reads handwritten digits, and learn the PyTorch workflow every ML team uses.
Uses: Deep learning in PyTorch, Neural networks5 stepsFine-tune a language model with LoRA
Teach a real language model a new style by training less than 0.2% of its weights, on a laptop CPU.
Uses: Training and fine-tuning, Deep learning in PyTorch, Tokens and embeddings5 stepsPut your project online
Turn the notes search engine into a tested web API, package it with Docker, and give it a public address.
Uses: Production AI and MLOps, Search and retrieval6 steps