AI Engineering

What the book covers

The book has ten chapters and runs to about 500 pages:

1. Introduction to building AI applications with foundation models (use cases, planning, the AI engineering stack)
2. Understanding foundation models (training data, architecture, post-training, sampling)
3. Evaluation methodology
4. Evaluating AI systems
5. Prompt engineering
6. RAG and agents
7. Fine-tuning
8. Dataset engineering
9. Inference optimisation
10. AI engineering architecture and user feedback

Two full chapters on evaluation come before prompting and fine-tuning, which reflects its central argument: without a reliable way to measure quality, every later decision is guesswork.

What makes it different

Knowledge about building with LLMs is scattered across blog posts, papers and threads, each covering one piece and assuming the rest. This book puts it into one coherent framework and, for each topic, gives you the questions to ask: should this be built at all, how will it be evaluated, when is fine-tuning worth it and when is it not, how much data is needed, how to make it faster and cheaper.

It is deliberately tool-agnostic. The author states it is not a tutorial book and has few code snippets, because tools date quickly and fundamentals last longer. Case studies and extensive references back the explanations.

The author's companion GitHub repository (chapter summaries, study notes, resources and case studies) had about 17,500 stars as of September 2026, and the book has been translated into more than a dozen languages.

Before you start

You should be comfortable with programming and have at least used an LLM API or built a small AI feature. Some basic machine learning vocabulary helps in chapters 2, 7 and 9, and the author flags the more technical sections so you can skim them on a first read.

How to read it

• Read chapters 1, 3 and 4 first, even if you are keen to get to RAG. They change how you approach everything else.
• Then read the chapter that matches your current problem: prompting, RAG and agents, or cost and latency.
• Keep a real project alongside. After each chapter, write down one change to make to your system and try it.
• Use the companion repository's chapter summaries for revision.

At a chapter a week, it takes two to three months.

How it fits with other resources

• Designing Machine Learning Systems (the same author) covers traditional ML systems: tabular data, feature engineering and model training. This book covers foundation models. Each stands alone, and many engineers read both.
• Hands-on courses and provider documentation teach the specific APIs; this book teaches how to decide what to build.

Common questions

Is it outdated already? Specific model names and benchmarks in it will date, but the evaluation, data and architecture fundamentals it focuses on are exactly the parts that last.

Is it for beginners? It is for technical readers. A student with programming experience can follow it, but it helps to have built something first.

Does it teach me to train an LLM? No. It is about adapting and building on existing foundation models.

Best for: engineers building AI products who want one structured, evidence-backed framework for the whole process.