A free 100-topic AI engineering roadmap you can run without an API key

A study order instead of twelve tabs

As dev.to reported, a developer got tired of opening a dozen browser tabs every time they tried to learn retrieval-augmented generation (RAG), agents or MCP, with no sense of what to read first. So they wrote down a sequence and then wrote the lessons to match. The result is a public code repository called ai-engineer-roadmap.

RAG is a technique where an AI system looks up relevant passages from a set of documents before it answers, so its reply rests on real text. Agents are AI programs that take steps toward a goal, such as calling tools. MCP, the Model Context Protocol, is a standard way for agents to connect to those tools.

The author says anyone who knows basic Python can start right away.

What the 100 topics look like

The material is split into seven phases:

• RAG Foundations: 20 topics
• Advanced RAG: 20 topics
• LLM Concepts: 14 topics
• Agentic AI: 16 topics
• Frameworks: 8 topics
• MCP: 7 topics
• Production AI: 15 topics

Every topic follows one template. It opens with a plain explanation built around a familiar comparison, then adds a diagram and a Python example. It ends with three interview questions and answers, plus links to the original research paper or the official documentation.

Topics are tagged beginner, intermediate or advanced. A reader with some background can skip ahead instead of reading in strict order.

Trying it without an account

The author describes a five-minute setup. You copy the repository to your computer, create an isolated Python environment, install the listed requirements, run the included tests, and then start a demo script called mini_rag.py with a sample question.

That script is a complete but tiny RAG pipeline. It splits three sample documents into pieces, scores each piece against your question, and prints the best match along with the file it came from.

The demo scripts rely only on tools that ship with Python, so they need no API key and no internet connection. There is one optional upgrade: if you set an Anthropic key and install the anthropic package, Claude writes the final answer from the retrieved passages.

Where the demos stop

The author is open about the limits. The mini RAG demo ranks text with simple word matching rather than a real embedding model, which is the kind of AI component that turns text into numbers so meaning can be compared. It is meant to show the shape of the pipeline, not to serve as a retriever you would deploy.

The code inside the topic pages is different. It uses real libraries such as sentence-transformers, chromadb and langgraph. Those must be installed, and some need an API key, though each is labeled.

The final phase includes Dockerfile and Kubernetes configuration examples. The runnable demos there use small stand-ins that work inside a single program, so reading along does not require Docker or Redis.

Checks, a correction and how to judge it

The author says automated checks confirm that links between pages point to real files, that the diagram files are valid, that the tests pass and that the TypeScript MCP example builds. They also fixed one of their own claims: an earlier version advertised 47 interview questions, but the real count is 43 plus 4 system-design prompts.

Those checks show the repo is internally consistent. They do not show the lessons are accurate or well paced. Our newsletter did not review the topics, and this account comes from the author's own post.

A sensible approach is to skim the topic list, mark what you already know, and choose the first three gaps rather than tackling all 100. Check one lesson against another source before you rely on it. Pair each topic with a small project of your own, since a reading list is not proof of skill.