// Reading Plan
The order, the focus, and the practice.
27 books across the AI engineering stack and the forward-deployed foundations, ordered by what compounds. Open any book for the chapters that earn their pages, what to build while reading, and the closed-book questions that prove it landed.
27 books · ~436 h reading + ~182 h consolidation = ~618 h total · core path ~391 h
// Why these books
The certificates answer: Have you been exposed to this? The books answer: Can you operate when nobody has given you the steps?
Completion is a rate. Depth is a stock.
A certificate is built to be finished: short lessons, guided labs, known answers. A book is the author's full argument, including the parts that do not fit in a lesson — why the happy path fails, what it costs, when not to use it.
The field moves faster than curricula.
A course from 2024 teaches 2024's tooling. The load-bearing ideas here — evaluation, the data lifecycle, architecture, cost — move far more slowly than the SDKs.
Certificates teach the vocabulary. Books teach the judgement.
The hard part of AI engineering is not calling a model; it is deciding what to build, how to know whether it works, and what to do when it drifts. That is judgement, built from depth.
A book is a fountain: the water is there whether you drink or not. Passive reading, highlighting, re-reading, and binge-reading all produce the memory of having read rather than the ability to use it. The difference between a reader who keeps the ideas and one who keeps only the memory is not talent — it is whether the reading demanded something of them.
// How to read them
The protocol is built on the learning-science literature — retrieval practice, spacing, interleaving, consolidation — not on reading harder.
Preview
Headings, summary, figures. Write one question the chapter should answer. Reading with a question is reading with a destination.
Read actively
Pen in hand, not highlighter. The claim in your own words, the evidence, what would make it false. Sketch the diagram — a second retrieval route.
Retrieve
Close the book. Write the chapter's argument from memory: what it was about, the key claims, the evidence, what is missing. This feels like wasted time. It is the step that does the work.
Space the review
Return to the recall note at 24 hours, 1 week, and 1 month. Each review is a retrieval, not a re-read.
Apply it
Build the smallest thing that uses the idea, or teach it. Knowledge that is used transfers; knowledge that is not used decays.
Sleep on it
Consolidation happens overnight. Trading sleep for reading borrows from the retention the reading depends on.
The standard
One session = one chapter = one artifact. If a session produced nothing you can point to — a recall note, a sketch, a piece of code — it did not happen.
- —No highlighting as a substitute for retrieval.
- —No re-reading before attempting recall.
- —No reading without a purpose.
- —No massed binge-reading.
- —No reading without application.
Retrieval practice
Recalling information strengthens it more than restudying it. Retrieval is a memory modifier, not a read-out.
Spacing
A review at the edge of forgetting reconsolidates the trace, stronger than before.
Interleaving
Mixing topics forces you to choose the right strategy, not just execute a known one.
Desirable difficulties
Conditions that feel harder in the moment — retrieval, spacing, interleaving — produce the durable learning. If it feels easy, it is probably not sticking.
Stage 1 — AI & ML Engineering Core
1,442 pp · ~60 hRead fully, in order. The mental model.Foundation models as components, and how to design around what they get wrong.
Stage 2 — LLM Applications & Operations
2,137 pp · ~90 hRead fully, after Stage 1.Build LLM applications, then run them in production.
Stage 3 — Architecture & Data Platforms
1,067 pp · ~46 hRead fully, after Stage 2.The shapes that outlive a framework: how a system is structured, split, and documented.
Stage 4 — Craft & Quality
1,438 pp · ~61 hRead fully, after Stage 3.Better code and better verification, opened when a problem asks for it.
Stage 5 — Reference
861 pp · ~36 hReference shelf. Read on demand.The algorithms, kept on the shelf for when you need the derivation rather than the recipe.
Stage 6 — Forward Deployed: People
3,259 pp · ~134 hRead fully. The human side of the stack.Know what the technology is for, find out what people actually need, and keep executing when the problem is ambiguous.
Domain Elective — Fraud & Risk
220 pp · ~9 hElective. Read when a fraud or risk problem appears.A vertical: how AI and agentic systems change fraud detection and investigation.
Cover art hotlinked from Open Library and Google Books; the one title with no catalogue art gets a typographic tile.
// Knowledge stress test
Reading is not learning. The test is whether you can reproduce the idea with the book closed.
- 1
Close the book
After each focus chapter, write its core idea in two sentences from memory. If you cannot, you read it but did not take it.
- 2
Explain it aloud
Teach the idea to an imaginary junior engineer. Where you hedge or reach for the book's exact phrasing, that is the gap.
- 3
Rebuild it by hand
Implement the mechanism in a scratch file — not a summary, a working version. Ten lines that run beat a page of notes.
- 4
Revisit on a schedule
Return at one day, one week, and one month. Each revisit should be faster. If it is not, the idea never landed.
Stage stress tests
- Stage 1Design an ML system end to end on paper — data, features, model, deployment, monitoring — then build the evaluation harness for it.
- Stage 2Ship a model behind an API with monitoring, offline evaluation, and a rollback path. The stage is done when it survives a week of traffic.
- Stage 3Refactor one Python service to the patterns in the book, and profile one hot path until the bottleneck moves.
- Stage 4Spend a week embedded with one user of a system you built: interview them, watch them work, and ship one change that removes a real obstacle. Then write down what you believed before and what the evidence changed.