// lessons

The foundations. What a neural network actually is, why stacking ReLUs can approximate any function, and how gradient descent tunes millions of parameters. Start here.

Theory into practice. Build a working neural network from scratch in a spreadsheet — no code, no libraries, just arithmetic you can see and touch. Train it on real Titanic passenger data.

From models to agents. What separates a chatbot from an agent, the six building blocks every production coding agent shares, how a single request flows from keystroke to output, and how permission systems keep an autonomous loop safe.

The machine in the middle. How a language model turns a transcript into the next token: tokens and embeddings, why attention was the breakthrough, what a context window physically is — and a complete tiny GPT in PyTorch you can read in one sitting.

The hands-on payoff. Take the TinyGPT from Lesson 4, feed it a laptop-sized corpus, write the training loop yourself, and watch generated text condense out of randomness — from noise, to word-shaped noise, to fluent (and confidently wrong) sentences. Complete runnable script included.

The shaping pass. How a raw next-token predictor becomes an assistant: instruction tuning, preference-based refinement, why the same weights can host such different behaviours — and how the post-training pipeline explains sycophancy, polished hallucination, and miscalibrated refusals in deployed clinical tools.

Grounding the model. How text becomes coordinates, why nearest-neighbour search finds meaning rather than keywords, the full RAG pipeline in twenty lines — and the retrieval failure modes (superseded guidance, similar-but-opposite passages) that matter most in clinical deployments.

Everything, assembled. Build a working retrieval assistant over a folder of your own documents — local embeddings, the twenty-line retriever, grounded answers with citations — then measure it honestly with a recall@k evaluation harness. Complete runnable script included.