A first-principles curriculum

The AI Engineer's Learning Path

A first principles,decision-driven curriculum that takes you from zero to production-ready in building with LLMs and generative AI.

8
Modules
34
Lessons
17
References
6
Provider Paths
Start with the Guides Browse Modules
How this learning path works

Start with the problem.
Experience the gap.
Then build the solution.

First Principles, Not Recipes
Recipes, cookbooks and playbooks help when you have a foundation to plug them into. Build that foundation here, learn core concepts, discover the reasoning behind techniques, and the mechanics that make them work.
Measure Before You Build
You can't improve what you don't measure. Define what "good" looks like before writing any agent code, and save the vibes for toy projects.
Vendor-Agnostic Learning
Learn the principles, use the frameworks. OpenAI, Gemini, Anthropic, Hugging Face, Ollama, and GitHub Models are all supported paths. The concepts matter more than any specific implementation.
Build a System, Not a Demo
Each module in this course builds on top of the previous, leading you to an agent build that can be used in production and adapted to new use cases.
Curriculum

8 tailored modules. What you'll learn in this path.

These topics are intentionally ordered so we cut through the hype and focus on what matters, building on something you've already experienced. You'll learn things because you need to know them, and how to discern whether or not they're needed.

Module 00
Orientation
Before diving into any modules, get familiar with the landscape. These guides cover what AI engineering actually is, how this path is structured, how to choose a provider, and the hard rules that apply throughout.
9 guides
0.1 About This Path 0.2 What AI Engineering Is (and Isn't) 0.3 Why This Order 0.4 How to Use the Path 0.5 Concept Map 0.6 Common Category Mistakes 0.7 Choosing a Provider 0.8 Choosing Your Track 0.9 Hard Rules
Module 01
Foundations of AI Engineering
Get oriented with the core concepts that underpin everything else in this path. Understand the landscape, the terminology, and the mental models you'll use throughout. This is the foundation the rest of the modules assume you have.
6 lessons
1.1 Python and FastAPI 1.2 LLM Mental Models 1.3 Prompt Engineering 1.4 Building with APIs 1.5 Retrieval Basics 1.6 Security Basics
Module 02
Benchmark and Harness
Define what "good" looks like before you write any code. Learn to measure outcomes, design benchmarks that matter, and establish baselines you can actually improve against. Leave the vibes to toy projects.
3 lessons
2.1 Choosing a Repo and Defining 'Good' 2.2 Benchmark Questions 2.3 Run Logs and Baseline
Module 03
Agent and Tool Building
Build a tool-calling agent from scratch, then rebuild it with a framework so you understand what the framework actually does for you. Learn to make your tools portable with MCP so they aren't locked to a single runtime. The raw loop comes first because you need to know what you're abstracting away.
4 lessons
3.1 Building a Raw Tool Loop 3.2 Rebuilding with a Framework 3.3 Building an MCP Server 3.4 MCP Resources and Prompts
Module 04
Code Retrieval
Learn retrieval by building it four different ways. Each approach handles something the last one doesn't, from naive vector search to full context compilation. The goal isn't to always reach for the most complex option, but to know when each one is the right call.
5 lessons
4.1 Choosing a Retrieval Strategy 4.2 Naive Vector Baseline 4.3 AST and Symbol Retrieval 4.4 Graph/Hybrid Retrieval 4.5 Context Compilation
Module 05
RAG and Grounded Answers
Retrieval on its own just finds things. This module connects retrieval to answer generation so your system can cite its sources and stay grounded in evidence. Learn to build context packs, route between retrieval modes, and know when the model has enough information to answer.
3 lessons
5.1 RAG Pipeline 5.2 Evidence and Context Packs 5.3 Retrieval Routing
Module 06
Observability and Evals
If you can't see what your system is doing, you're just vibing. Build telemetry into your agent from the start, track costs and outcomes, then write evals that hold it accountable. This module ties measurement back to the benchmarks you built in Module 02.
5 lessons
6.1 Telemetry 6.2 Cost, Caching and Rate Limits 6.3 Building Your AI Harness 6.4 Retrieval and Answer Evals 6.5 Tool-Use and Trace Evals
Module 07
Orchestration and Memory
Learn to coordinate multiple agents, design specialists that handle specific tasks, and route work between them. Then tackle memory, both short-lived thread state and long-term persistence, so your system can remember what matters across conversations.
5 lessons
7.1 Orchestration 7.2 Specialists and Routers 7.3 Agent-to-Agent Interop 7.4 Thread and Workflow Memory 7.5 Long-Term Memory and Write Policies
Module 08
Optimization
Premature optimization is a costly distraction. Until your system is stable and measured, tuning it is guesswork. Learn when distillation and fine-tuning actually make sense, and how to apply them without breaking what already works.
3 lessons
8.1 Optimization Ladder 8.2 Distillation 8.3 Fine-Tuning
Reference

References to keep within reach

Lookup tables, schemas, decision frameworks, and terminology you'll return to throughout the path.

Pinned
Glossary Decision Rules Debugging Heuristics Excluded Topics and Clarified Terms
All references
Retrieval Substrate Chooser Run-Log Schema Telemetry Event Model Eval Taxonomy Context-Pack Contract Suggested Repository Layout Cost and Reliability Patterns Hardware and Model Size Guide Model Selection and Serving Model-Provider Matrix Topical Exercise Bank Portfolio Milestones Cloud Provider Surfaces (Bedrock, Vertex AI)
Provider Agnostic

Every example runs on your provider

Direct provider APIs
OpenAI
The most widely documented API surface. If you already have an OpenAI platform key, start here.
Gemini
Google's direct model API. A strong free tier and native SDK make it a practical alternative to OpenAI.
Anthropic
Claude's direct API, or through AWS Bedrock. Strong structured output and tool-calling support.
Also supported through
Hugging Face
Routes to thousands of open models through one free account. A good path if you want to learn without spending.
Ollama
Run models locally or through Ollama Cloud. The best option for privacy, offline work, or predictable costs.
GitHub Models
Hosted inference through your GitHub account. Access multiple model publishers from a single API surface.

These are just the surfaces we interact with. The concepts are the focus, and your learning isn't gated behind any single vendor.

Get started

Get started orienting yourself to the AI space before starting any modules

Start with the guides to understand the approach, choose your provider, and set up your environment. Then Module 1 begins.

Start with the Guides