AI Engineer Roadmap
The full map: models, RAG, context, agents, MCP, evals. Deep, but always grounded in practice.
Intermediate · 33 lessons · ~7.5 h
Lessons
Start here
- 1What is an AI Engineer?The first step of the AI Engineer roadmap: what the role actually is, how it differs from ML engineers and researchers, and why it exists at all.
- 2How LLMs Actually WorkWhat a Large Language Model really is: next-token prediction, tokens, training vs inference, the context window — and why models sometimes confidently make things up.
- 3AI vs AGI: Where We Actually AreNarrow AI is what we have, AGI is what we don't. What the difference means in practice, why today's models are tools rather than agents-of-the-self, and why builders should care.
Models
- 4Types of AI ModelsPre-trained models, closed vs open weights, self-hosting. Who makes them — OpenAI, Anthropic, Google, Meta, Mistral, DeepSeek — and how to pick a lane.
- 5How to Choose the Right ModelCapability vs cost vs speed vs context size. A practical checklist for picking a model, and when a small cheap model is all you need.
- 6Talking to a Model: Tokens, Context and SamplingThe physics of a conversation with an LLM: how text becomes tokens, why the context window overflows, and what temperature, top-p and top-k actually do.
- 7APIs and SDKs: How Code Talks to ModelsThe OpenAI-compatible standard, the Claude Messages API, the Gemini API, and the Hugging Face ecosystem — with real code you can copy and run.
Prompting
- 8Prompt Anatomy: What a Prompt Is Made OfThe building blocks of a good prompt: system prompts, input format, constraints and examples — plus a checklist and a bad-vs-good prompt you can copy.
- 9Prompting Techniques: Zero-shot, Few-shot, CoT, ReActThe four core prompting techniques every AI engineer uses: zero-shot, few-shot, chain-of-thought and ReAct — with copyable prompts and when each one helps.
- 10AI Safety Basics for BuildersPractical AI safety: prompt injection with a live demo, bias, privacy and moderation — and the defenses you can ship this week. Not scary, just engineering.
Embeddings
- 11What Are Embeddings?The idea that powers semantic search and RAG: turning text into vectors of numbers so that similar meanings land close together. Plus where embeddings are used in real products.
- 12Embedding Models and ProvidersOpenAI, Cohere, Gemini, Jina, or open-source Sentence Transformers? Dimensions, cost, multilinguality — and a practical way to choose an embedding model.
- 13Vector DatabasesWhy semantic search needs a vector database, how ANN indexes like HNSW find nearest neighbors fast, which one to pick, and a working similarity-search example.
RAG
- 14What is RAGRetrieval-Augmented Generation: give the model your own data at answer time instead of hoping it memorized it. Why hallucinations happen, real use cases, and RAG vs fine-tuning.
- 15The RAG pipeline, step by stepChunk your documents, embed the chunks, index them, retrieve at query time, assemble the prompt, generate. Every stage of RAG with a walkthrough and the token math of a real prompt.
- 16Ways of implementing RAGRaw SDKs, LangChain, LlamaIndex, Haystack, or fully manual: the trade-offs of control, convenience, and lock-in — with a minimal RAG in ~30 lines of Python, no API keys.
Context
- 17Context vs Prompt Engineering: What the Model Actually SeesYour prompt is one message; the context is everything the model sees — system prompt, memory, retrieved docs, tool results, history. Why context engineering is becoming the bigger discipline.
- 18Context Engineering Techniques: Compaction, Isolation, RAG and MoreThe five core context engineering techniques — compaction, isolation, RAG with dynamic filters, long-context processing, and state management — and when to apply each.
- 19Context Failure Modes: When More Context Makes Things WorseLost-in-the-middle, context poisoning, context rot, and contradictory instructions — how context breaks, how to detect it, and how to prevent it.
Agents
- 20What Are AI Agents?A model in a loop with tools: think, act, observe, repeat. What agents are, how they differ from chatbots, and why they only became practical recently.
- 21Tools and Function CallingHow a model calls your code: tool schemas, the full function-calling round trip, a working weather bot, and the practices that keep agents reliable.
- 22Agent SDKs and FrameworksOpenAI Agents SDK, Claude Agent SDK, LangChain and LlamaIndex — what an SDK actually gives you, and when building the loop yourself is the better deal.
- 23Multi-Agent SystemsWhen one agent is not enough: context isolation, specialization, parallel work, the orchestrator-worker pattern, and the pitfalls nobody warns you about.
MCP
- 24MCP: The USB Port for AIThe Model Context Protocol — the open standard that connects models to tools and data, and why the industry needed it.
- 25Build Your First MCP ServerA step-by-step MCP server in Python with the official SDK: tools, resources, prompts, stdio transport, and wiring it into a client.
AI for coding
- 27AI-Assisted Coding: The New ToolsClaude Code, Cursor, Windsurf, Codex, Gemini, Replit and opencode. What agentic coding tools do differently from autocomplete, and how to choose one.
- 28Working with Coding AgentsHow to actually get good results from coding agents: context files, small tasks, reviewing diffs, tests as guardrails, and when not to delegate at all.
Quality: observability and evals
- 29LLM ObservabilityWhy AI features need monitoring beyond normal logs: tracing every request — prompt, model, tools, tokens, cost, latency — plus production monitoring and the tools that do it.
- 30LLM EvaluationsWhy 'it seems fine' is not a QA strategy: deterministic checks, LLM-as-judge and human evals — with a tiny working eval harness you can run in the browser.
- 31Regression Testing for PromptsPrompt changes break cases you never tested: golden datasets, running evals on every prompt or model change, and the CI gate that blocks bad deploys.