louisfb01/start-ai-engineering: a curated roadmap for learning AI engineering in 2026

A reading list, not a codebase

The repository has three top-level entries: README.md, an assets/ directory holding images, and people-social-tagging-list.md. There is no package manifest, no source directory, no build step. That matters because the name sounds like a starter kit and it is not one. The README states the guide is "intended for anyone with zero or a small background in programming, AI, or machine learning who wants to become a strong AI engineer in 2026," and that it is organized by how you like to learn: videos, articles, books, docs, courses, and real projects.

The practical consequence is that your first decision is not which branch to clone but which track to follow. The README says there is no single correct order and that a classic path runs top to bottom, with explicit permission to skip books if you dislike them or skip online courses if you do not want one. That is a deliberate design choice: the guide is a menu, and the maintainer expects you to leave parts of it untouched.

Who it is for, then, is someone who already knows they want to learn this field and needs a curated index. Who it is not for is someone who wants a template repository to fork, or an evaluator who wants to check a claim by running something. There is nothing here to run.

How the README is structured and how to read it

The table of contents is the real product. It runs from prerequisites and a suggested learning path through prompting, reasoning models and test-time compute, context engineering, RAG, embeddings and vector databases, tools and MCP, workflows and agents, evaluations and observability, fine-tuning, multimodal work, voice agents, deployment, AI coding agents, safety and guardrails, then communities, newsletters, people to follow, and a job-search section.

Two mechanisms are worth noting. First, the difficulty guide: resources carry compact markers from 1 to 10, where the README describes 1 as absolute beginner, 3 as beginner-friendly AI vocabulary, 5 as practical builder material, 7 as production engineering depth, and 9 or 10 as advanced systems, research, or senior-level papers. That scale is the closest thing the repository has to a dependency graph, and it is what lets you pick an entry point without reading every link. Second, the suggested learning path, which the README gives as a numbered sequence: watch a few foundational videos, pick one free course and one framework whose docs you read end to end, pick one or two books, optionally take advanced applied courses, build two or three small but real projects that break in interesting ways, then add evaluations, tracing, and deployment before calling anything production-ready.

The README is also opinionated about tooling in a way that shapes the reading. It warns that coding agents such as Codex, Claude Code and Cursor can scaffold apps and speed up steps, and that you should use them, but it frames AI engineering as "the judgment layer behind the work": deciding what to build, which architecture fits, how to evaluate it, where it will fail, and whether it is reliable enough to ship. That framing is why the sections on evaluations, observability and harnesses sit alongside the ones on prompting and RAG rather than after them.

Starting with the guide: no install, just a clone and a read

There is nothing to build. The repository is documentation, so the first real use is fetching the text and reading the table of contents against your own background. If you want a local copy to annotate or to diff over time, clone it:

git clone https://github.com/louisfb01/start-ai-engineering.git
cd start-ai-engineering

After that you should see README.md, the assets/ directory, and people-social-tagging-list.md. The README is the whole guide; the other two entries are supporting material, not entry points.

The README offers a second starting move that is specific to this project: it says you can personalize the roadmap with an AI agent, and that a prompt is provided in the README for that purpose. The visible text cuts off mid-sentence at "Paste the prompt bel", so the full prompt text is not shown here. If you want to use that route, read the live README rather than relying on a summary, because the prompt is the part that tells the agent what to do with your background and goals.

A reasonable first pass, based on the README's own path, is to pick one free course and one framework's documentation and commit to reading the docs end to end, then choose two or three project ideas from the practice section. The guide does not tell you which course or which framework to pick; that choice is left to you, and it is the decision that most affects whether the roadmap works.

Where the guide is thin, and where it will waste your time

The most obvious limitation is that a static Markdown file ages at the speed of the field it describes. The README acknowledges this directly: "This guide is updated throughout 2026 as the stack moves." The last push to the repository was on 2026-08-03, so the content is recent, but "updated throughout 2026" is a promise about cadence, not a guarantee about any individual link. Nothing in the repository verifies that a listed video still exists, that a course is still free, or that a framework's documentation still matches the version you install. You are the link checker.

A second limitation is the monetization layer. The README states that most resources are free, that paid resources are clearly labelled, and that some paid course and book links are affiliate links that support the guide at no extra cost to you. That is disclosed plainly, which is better than most curated lists manage, but it is still a structural bias: a paid course that pays a commission and a free course that does not are not competing on equal footing in the ordering. Read the labels, not just the list.

Third, the guide is explicitly not a substitute for building. The README's own path puts projects at step five, after videos, a course, books, and optional advanced courses. If you follow it literally and stop before the project step, you will have consumed a lot of material and shipped nothing. The README says as much: "Repetition and debugging are where the real learning happens." The wrong user for this repository is anyone hoping the reading list alone converts into engineering skill, and anyone who needs an auditable curriculum with assessments rather than a curated index.

How this differs from a structured course or a framework's own docs

The nearest alternative is a single structured course, and the difference is coverage versus sequence. A course gives you one author's ordering, exercises, and usually some form of completion signal. This guide gives you many authors' material across prompting, RAG, agents, evals, deployment and safety, with a difficulty marker on each entry and an explicit statement that there is no single correct order. If your problem is that you do not know what exists, the guide wins. If your problem is that you keep starting and stopping, a course with a fixed syllabus wins, because the guide deliberately refuses to enforce one.

A second alternative is the documentation set of a single framework, read end to end. The README actually recommends this as part of its path, which makes the two complementary rather than opposed: the guide tells you which topics exist and roughly how hard they are, and the framework docs tell you how one concrete implementation works. The failure mode of docs-only learning is that you learn one tool's vocabulary and mistake it for the field's. The failure mode of guide-only learning is that you accumulate vocabulary with no implementation attached.

A third alternative, for people who already write software, is to skip the introductory tiers entirely and work from the production-oriented sections. The difficulty markers make that possible: the README places practical builder material around 5 and production engineering depth around 7, so a working developer can start higher and use the lower tiers only as a gap check. That is a genuine advantage of the marker system over a linear course.

Maintenance, licensing and what the repository does not state

On maintenance, the facts are narrow. The repository is not archived, and the last push was on 2026-08-03. There are no retrieved releases, which is consistent with a documentation-only project: there is no versioned artifact to upgrade, so "upgrade cost" here means re-reading the README and re-checking the links you depend on, not migrating an API. The README describes the guide as updated throughout 2026 as the stack moves, and names louisfb01 as the maintainer, with contributions invited through pull requests.

On licensing, the repository does not state a license, and no license file appears among the top-level entries (README.md, assets/, people-social-tagging-list.md). That is worth flagging before you reuse the text, mirror the list, or republish it in a course of your own: without an explicit license, the default position is that no reuse rights are granted, and you should ask the maintainer rather than assume. This is not legal advice; it is a statement about what the repository does and does not say.

One more maintenance consideration is specific to a link list. The affiliate disclosure is part of the README, so if the disclosure changes, the meaning of the list changes with it. Anyone treating this repository as a stable reference should pin a commit rather than track main, so that a later edit does not silently change what you cited.

Editorial conclusion

Adopt this guide if you are starting from zero or moving into AI engineering from another discipline and you want a single index of free videos, courses, books and project ideas instead of a scattered bookmark folder. Skip it if you need runnable code, a framework, or a certificate; the repository ships no software and no credential. Before you commit, open the README and confirm three things: that the sections you care about (RAG, agents, evals, deployment) still list resources you can reach, that the paid and affiliate links are labelled as the README says they are, and that the difficulty markers 1 to 10 actually match your level on the first two or three entries you pick.

Frequently asked questions

Official sources

1. Issues
2. louisfb01/start-ai-engineering on GitHub
3. README