Learn AI by Building: The Free Resources I Made for Students, Engineers and Teachers

This month I sat down to prepare a 60-minute training on Retrieval-Augmented Generation. By the end of the week it had grown into something larger: a guide, an interactive workshop, a hands-on lab and a notebook a student could take home. Slides explain an idea. Building it is how the idea stays.

There is a Sanskrit verse I return to often: vyaye kṛte vardhata eva nityaṃ, vidyādhanaṃ sarvadhana pradhānam. Every other wealth shrinks when you spend it; knowledge grows each time it is given away. So everything below is free, open to anyone, and meant to be shared.

Why hands-on, and why one example

Most AI learning material stops at definitions. You read what an embedding is, nod, and forget it by Friday. The gap is not information; it is practice. You understand chunking the moment you watch a deadline get separated from its exception. You understand access control the moment a billing clerk is refused a confidential rate and a contract manager is not.

So every resource uses the same running story from my own domain, US healthcare revenue cycle: a payer denies a claim, and staff need to know the appeal deadline and what to attach. The policies change over time, differ by payer, and some are restricted. It is exactly the kind of problem where a confident but ungrounded AI answer causes real harm, which makes it a good teacher. (The payers and figures are illustrative.)

The learning path: start wherever you are

Step	Resource	What you get	Time
1	AI Concepts	A map of modern AI in seven layers, from foundations to agents and safety, with plain-language definitions.	20 min
2	RAG Explained	A basic-to-advanced guide: chunking, embeddings, hybrid search, reranking, filters, agentic RAG, GraphRAG, evaluation, security and production.	45–60 min
3	RAG Blueprint Workshop	An interactive 60-minute session with live labs, quizzes, a session clock and facilitator notes for anyone who wants to teach it.	60 min
4	RAG Lab	Build a small RAG app yourself in 16 runnable steps. The Python runs inside your browser: no sign-in, nothing to install.	60–90 min
5	The full notebook in Google Colab	The complete version with a neural embedding model, a reranker and optional answers from Claude. You get your own copy.	60–90 min
6	Harness Engineering (GitHub)	My open-source kit for making AI coding agents checkable: proof matrices, verification gates, session handoffs and runnable demos.	Self-paced

By the end of step 4 you will have written an ask() function that finds the right policy, respects who is asking and when, answers with citations, and says “I don’t have enough evidence” instead of guessing. You will also have measured it, which is the habit that separates a demo from a system.

Short reads to go with them

• From Confused to Confident: the AI Glossary, for the vocabulary everyone uses and few explain.
• Build a Python Coding Agent in 30 Minutes, if you want to see how agents work from the inside.
• Engineering & AI, where all my essays on AI adoption, agents and engineering leadership are collected in one place.

How different people can use them

• Students: follow steps 1 to 4 in order, then try the final challenge in the lab: replace the sample library with notes from a subject you know, write five test questions first, and measure.
• Engineers: go straight to the lab and spend your time on filters, the evidence gate and evaluation. That is where production systems succeed or fail.
• Teachers and trainers: run the workshop in class with the facilitator notes switched on, and give the notebook as homework. You are welcome to use these with your students; a mention of where they came from is appreciated.
• Leaders: read AI Concepts and the production module of the workshop. You will come away with better questions to ask your teams than “which model are we using?”

The one lesson underneath all of it

After 25 years of building software for hospitals, the pattern I see most often is this: when an AI system gives a wrong answer, the model is rarely the real culprit. The cause sits in the data, the retrieval, the permissions or the missing check. A model is only as good as the evidence it is given. These resources are designed so you can see that with your own eyes, and fix it with your own hands.

Knowledge is the only wealth that grows when it is spent. Spend it generously.