How the AI Engineering from Scratch Curriculum Is Structured: 20 Phases, 511 Lessons, and 3 Learning Paths

The AI Engineering from Scratch curriculum uses a layered architecture of 20 sequential phases containing 511 lessons, each following a 5-step "build-it-then-use-it" pattern to progress from basic tooling to production AI systems.

The rohitg00/ai-engineering-from-scratch repository implements one of the most comprehensive open-source curricula for learning AI engineering. Unlike fragmented tutorial collections, this curriculum enforces strict structural conventions across every lesson, enabling learners to generate reusable artifacts—prompts, skills, agents, and MCP servers—that compound across phases.

Phase-Based Architecture: 20 Levels of Progression

The curriculum organizes content into 20 phases, visualized in the README.md Mermaid diagram spanning lines 78–99. Each phase represents a broad topical block that stacks on previous knowledge:

Phase Range	Focus Area	Example Phase
0–2	Tooling and mathematics	phases/01-math-foundations/
3–12	Core ML and deep learning	phases/09-transformers/
13–18	Tools, protocols, and agents	phases/14-agent-engineering/
19	Integration and capstones	phases/19-capstone-projects/

The phase map in README.md explicitly shows dependencies, allowing learners to skip ahead when they master lower layers while preserving logical connections that explain why higher-level concepts work.

Lesson Structure: 5 Beats and 4 Output Types

Every lesson across all 511 units follows an identical folder layout in phases/<NN>-<phase-name>/<NN>-<lesson-name>/:

<lesson>/
├── code/          # Runnable implementations

├── docs/          # Narrative content (docs/en.md)

└── outputs/       # Generated artifacts

As documented in README.md lines 110–118 and 124–132, each lesson implements five beats:

1. MOTTO – Core principle in one sentence
2. PROBLEM – Concrete challenge to solve
3. CONCEPT – Theoretical foundation
4. BUILD IT – Implementation phase
5. USE IT → SHIP IT – Deployment and real-world application

This structure ensures every lesson produces a reusable artifact—either a prompt, skill, agent, or MCP server—that learners carry forward into subsequent phases.

Code Implementations: Up to 4 Languages Per Lesson

Lessons provide runnable implementations across multiple languages. For example, Phase 1's linear algebra intuition lesson:

# Clone and run a sample lesson

git clone https://github.com/rohitg00/ai-engineering-from-scratch.git
cd ai-engineering-from-scratch
python3 phases/01-math-foundations/01-linear-algebra-intuition/code/vectors.py

The code/ directory in each lesson contains language-specific implementations, with Python as the primary reference and additional languages where pedagogically relevant.

Three Entry Points: Adaptive Learning Paths

The curriculum supports three distinct entry points, defined in the skills/ directory and documented in README.md lines 64–71:

Entry Point	Command	Generated File	Best For
Placement tutor	learn start-learning	LEARNING.md	Assessing current skill level
MCP path	MCP host integration	MCP-LEARNING.md	Engineers building tool ecosystems
Agent-Skills path	Agent host integration	AGENT-SKILLS-LEARNING.md	Developers creating autonomous agents

To use the placement tutor:

# Requires npx and a skill-capable host

npx skills add rohitg00/ai-engineering-from-scratch
learn start-learning

Each path generates a personalized learning file that guides the learner through a relevant subset of phases rather than enforcing linear progression.

Capstone Projects and Certification Tracks

Phase 19 aggregates capstone projects that combine artifacts from earlier phases into end-to-end systems. Located in phases/19-capstone-projects/, these projects require integrating prompts, skills, agents, and MCP servers built throughout the curriculum.

Separate Claude certification tracks in certifications/claude/ map lesson subsets to official exam objectives. These tracks reuse the same lesson infrastructure while adding diagnostic assessments aligned with certification requirements.

MCP Labs and Protocol Implementation

Phase 13 covers the Model Context Protocol (MCP), with lessons like phases/13-tools-and-protocols/06-mcp-fundamentals/ demonstrating protocol implementation:

# Run an MCP fundamentals lab

python3 phases/13-tools-and-protocols/06-mcp-fundamentals/code/run_mcp.py

The skills/learn-mcp/SKILL.md file defines the MCP learning path, enabling integration with MCP-compatible hosts for hands-on protocol practice.

Summary

• 20 phases provide topical progression from tooling to production systems, visualized in README.md with explicit dependency graphs.

• 511 lessons follow identical 5-beat structure (MOTTO → PROBLEM → CONCEPT → BUILD IT → USE IT → SHIP IT) with consistent folder layouts.

• Four artifact types—prompts, skills, agents, and MCP servers—accumulate across phases and compose into capstone projects.

• Three adaptive entry points via the skills/ directory generate personalized learning files based on learner background and goals.

• Phase 19 capstones and Claude certification tracks provide integration testing and credentialing pathways using the same lesson infrastructure.

Frequently Asked Questions

How many lessons are in the AI Engineering from Scratch curriculum?

The curriculum contains 511 lessons distributed across 20 phases. Each lesson follows the same structural conventions with code/, docs/, and outputs/ directories, as defined in README.md lines 110–118.

Can I skip phases if I already know the material?

Yes. The phase architecture explicitly supports skipping— the README.md dependency graph (lines 75–77) shows logical connections between phases, allowing learners to enter at appropriate skill levels while maintaining understanding of why higher-level concepts work.

What programming languages does the curriculum use?

Lessons primarily use Python for implementations, with up to four languages available in select lessons. The code/ directory in each lesson contains runnable examples, with Python serving as the reference implementation across all 511 lessons.

How do the certification tracks differ from the main curriculum?

The Claude certification tracks in certifications/claude/ map a subset of lessons and diagnostics to official exam objectives. They reuse the same lesson infrastructure as the main curriculum but add assessment-focused scaffolding for credentialing purposes.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}