AI engineering

Most engineers will never train a model. Almost every engineer will build on top of one. This path is for that job: putting language models inside real products, where answers have to be right, actions have to be safe, and the bill has to make sense.

It is not a machine learning course. There is no calculus here and no training loop. It is system design with a model in the middle: how the model actually behaves, how to give it the facts it does not have, how to let it take steps without letting it take liberties, how it reaches tools and other agents, and how to know whether any of it works. Code examples come in both Python and TypeScript, because both are first class for this work now. Every chapter has something to play with, and every one ends with the version you would say out loud in an interview.

Start here#

What a model actually does with your text, what it costs, and how to talk to it so code can rely on the answer.

Retrieval#

Giving a model the facts it was never trained on, and finding the right ones.

Advanced RAG#

The techniques production systems use once basic retrieval stops being enough: better queries, better indexes, retrieval inside an agent loop, graphs, documents full of tables and images, and measuring all of it.

Agents#

Models that take steps: when to let them, how to design the loop around them, the frameworks people build them with, and how to keep a person in charge of what matters.

MCP and agent protocols#

How agents reach tools and each other: designing tools, the Model Context Protocol as of its 2026 revision, building and securing servers, and A2A for agent to agent work.

Production#

Knowing it works, keeping it safe, and keeping it affordable when people actually use it.

Design questions#

The full interview format: clarify, estimate, draw, go deep, and defend the trade-offs.