The Complete AI Engineering Curriculum
This is the full syllabus for going from zero to production as an AI engineer — a free, hands-on, learn-by-doing roadmap of 38 tracks and 231 lessons. Every lesson has objectives, a lab you type out yourself, runnable code, exercises with solutions, and common-mistake warnings. Below, the whole path is grouped into six phases, with a real link into each track so you always know exactly where to go next.
Who this is for & what you'll be able to do
- Who it's for: career-changers and developers who can write a little Python and want a single, ordered path to building real AI systems — not a link dump.
- Build a RAG system, a tool-using agent, and multi-agent teams from scratch, then harden them for production.
- Evaluate quality with golden sets and LLM-judge evals so you can prove a change helped before you ship it.
- Operate a live LLM app: caching, observability, guardrails, cost control, and deployment on your own infra or AWS.
- Ship a résumé-worthy capstone — an AI DevOps Engineer — plus a gallery of 15 portfolio projects.
The curriculum outline
The 38 tracks grouped into six logical phases. Each entry links to that track's first lesson or landing page — start at the top of a phase and work down.
Phase 1Foundations — the engineering bedrock
Before any AI, get fluent in the tools every engineering job assumes. New to Python? This phase alone is a full computer-science grounding.
🧰 Getting Started
Install VS Code and the Python + Anthropic extensions, run the labs, and keep your API key safe — the one-time setup every chapter assumes.
🐍 Python Fundamentals
Basics to expert on one umbrella page: types, data & structures, functions/OOP, and the advanced patterns agents are built from.
🧮 Data Structures & Algorithms
Complexity, stacks/queues/lists, hashing, trees/heaps, graphs, sorting — a full DSA track in Python, up to Big Tech patterns.
📐 System Design
SQL & schema design, OOD/SOLID & patterns, HLD & distributed architecture, and scalable infrastructure — taught Python-first.
🧑💻 Developer Foundations
The command line, Git & GitHub (branches, PRs, conflicts), and how the web works (HTTP, DNS, REST, JSON).
✅ Software Testing
pytest, unit & integration tests, TDD, fixtures and mocking — the confidence to change code without fear.
🐳 Containers & Deployment
Docker from scratch, images & layers, docker-compose, deploying to the cloud, and a CI/CD pipeline.
⚙️ Advanced AI Engineering
The production Python layer: metaprogramming, memory/GIL, async/streaming, tensors, tokenization, validation, vector DBs and MLOps.
Core LLM, RAG & Agents — the heart of the course
The essential arc: understand agents, make real API calls, ground the model in data, give it tools, then measure and harden it. This is the destination for most learners.
🧠 Foundations & Concepts
The mental model with no code: chatbot vs agent, the building blocks, the agent loop — plus the learning roadmap.
💬 Prompting & LLM Basics
Your first streaming API call, then designing system prompts, few-shot examples, and schema-valid structured output.
🧲 Retrieval (RAG)
Build RAG from scratch: chunk, embed, store vectors, retrieve with hybrid search + re-ranking, and cite grounded answers.
🤖 Agents & Tools
Define tools, write the agentic loop by hand, add human-in-the-loop gates, and give the agent memory across turns.
🛡️ Evaluation & Production
Golden sets, deterministic + LLM-judge evals and a regression gate, then caching, observability and deployment.
🏗️ FDE & Capstone
The Forward Deployed Engineer method, then assembling everything into the capstone AI DevOps Engineer, step by step.
Frameworks & Providers — build faster, deliberately
The industry-standard frameworks and provider skills on top of the hand-written loop — always mapped back to what you already built.
🎓 Specialized Topics
The gaps most courses skip: LLM security, multimodal AI, MCP, the fine-tune-vs-RAG-vs-prompt decision, and SQL.
🅰️ Claude & Anthropic
Working directly with the models: picking a tier, the Anthropic API and tool-use loop, Claude Code, and MCP.
🧰 Anthropic API in Practice
The day-to-day API as runnable recipes: Message Batches, prompt caching, Files/vision, token usage and server-side tools.
🅰️ Anthropic Skills
Claude Code in action & automation, Cowork, Agent Skills, subagents, advanced MCP, and Claude on Vertex/GCP.
🧩 No-Code Agentic AI
Ship the same agents on a canvas: n8n, Zapier, Make and Flowise — with each node mapped to the concept it implements.
⛓️ LangChain & LangGraph
Named architectures, LangChain Core (chains, memory, RAG), and LangGraph for stateful, cyclic, resumable agents.
👥 Multi-Agent Orchestration
CrewAI and AutoGen teams, agentic RAG & GraphRAG, and Deep Agents for long-horizon work.
Production & Ops — keep a live system trustworthy
Building the agent is half the job; keeping it healthy, cheap and trustworthy for a year is the other half.
📦 MLOps & LLMOps
Foundations & lifecycle, the infra/tooling stack, deployment & scaling, and monitoring/governance/responsible AI.
✨ AI-Assisted Development
Vibe-coding fundamentals plus Cursor, Google Antigravity, Amazon Q Developer and GitHub Copilot.
🔗 Interoperability & Agent Ops
MCP ecosystem integrations, the A2A/ACP/ANP protocols, tracing with LangSmith, and guardrail frameworks.
📊 Data & App Building
NumPy & Pandas, Matplotlib & Seaborn, AI backends with FastAPI, and front-ends with Streamlit & Gradio.
🎛️ Prompt & Context Engineering
Advanced prompting, context engineering, prompt optimization with DSPy, and working with LLMs & OSS models.
🔤 NLP & Transformers
NLP foundations, text representation & classification, neural sequence models, and transformers → LLMs.
☁️ AWS AI Automation
Bedrock (Claude, tool use, Knowledge Bases, Agents, Guardrails), SageMaker, and the AI services — all in code.
Specializations — go deep where it counts
Optional deep dives once the core is solid. Pick the ones that match your goals — frontier capabilities, systems internals, cost, tuning, self-hosting and safety.
🎙️ Frontier Agent Capabilities
Voice & realtime agents, computer use, reasoning models, long-term memory, and advanced retrieval & RAG evaluation.
🧬 ML Systems Internals
GPU memory, distributed training, attention & GPU kernels, multi-GPU serving, and throughput economics.
🧩 Advanced Challenges
Debug a broken RAG or agent, design under constraints, run an incident post-mortem, and defend a tradeoff.
⚡ Inference & Cost
Latency/throughput/cost, quantization, the KV-cache, batching & scheduling, speculative decoding and serving engines.
🎚️ Fine-tuning
When & why, dataset preparation, LoRA/QLoRA, full vs PEFT, DPO & preference tuning, and evaluate & serve.
🖥️ Local & Open Models
The open-weight landscape, local dev with Ollama, production serving with vLLM, quantized inference, and API vs self-host.
🚨 Safety & Red-teaming
Threat modeling, manual & automated red-teaming, layered defenses, and governance & compliance.
Career & Projects — prove it and get hired
Turn the curriculum into a portfolio and an offer. Build real projects, study reference architectures, and prepare for interviews.
🎯 Career & Interview Prep
Résumé & portfolio, behavioral/STAR, the coding & system-design interviews, and offers & negotiation.
🚀 Projects
Fifteen in-demand builds — support, code review, document intelligence, data analyst, deep research and more.
🏛️ Case Studies & Architectures
Agents vs workflows, support triage, compliance docs, healthcare RAG, scale war-stories, and constitutional AI.
How to use this curriculum
- Do the one-time setup (Getting Started) and read the Learning Roadmap for the mental model before writing any code.
- If you're new to Python, work Phase 1 top to bottom. Already fluent? Skim it and jump straight to Phase 2.
- Work Phase 2 in order — it's the spine of the whole course, ending in the capstone. Type every lab out yourself and run every snippet.
- Pull in Phase 3–4 as you need them: reach for a framework or a provider feature when a project demands it, not before.
- Treat Phase 5 as a buffet — pick the specializations that match your target role rather than doing all of them.
- Ship a Phase 6 project early and keep improving it. A finished, evaluated build beats ten half-read tracks.