ai-engineering-from-scratch

Product Introduction

1. Definition: AI Engineering from Scratch is an open-source, comprehensive curriculum and code repository for learning artificial intelligence and machine learning engineering. It is a technical educational resource, structured as a sequence of hands-on coding lessons.
2. Core Value Proposition: It exists to solve the problem of scattered, framework-first AI education by providing a unified, foundational spine. The curriculum ensures deep understanding by having learners implement every core algorithm—from linear algebra to transformer architectures and agent systems—using only raw mathematics and code, before introducing any high-level libraries like PyTorch or TensorFlow.

Main Features

1. Math-First, Library-Last Pedagogy: Every lesson requires deriving the underlying mathematics and then implementing it in code from scratch. How it works: Learners start with fundamental operations (e.g., matrix multiplication, gradient calculation) and build upward, only using frameworks like PyTorch after they have built a lower-level version themselves. This uses pure Python, NumPy for numerical computation, and later integrates industry-standard libraries for comparison.
2. Multi-Language Implementation (Python, TypeScript, Rust, Julia): The curriculum provides implementations in four programming languages, chosen based on the concept being taught. How it works: Core mathematical and neural network concepts are often taught in Python. Systems programming, performance-critical components, or server-side logic may be demonstrated in Rust or TypeScript, while Julia is used for scientific computing aspects. This exposes engineers to the practical ecosystem of AI tools.
3. Integrated AI Tutor & Skill System: The course can be consumed interactively via a command-line tutor. How it works: Using the command npx skills add rohitg00/ai-engineering-from-scratch, learners install a "skill" that allows an AI agent (compatible with Claude, Cursor, Codex) to act as a personalized tutor. This agent administers placement quizzes, recommends personalized learning paths, and teaches lessons interactively within the terminal.
4. Phase-Based Curriculum with Verifiable Artifacts: The learning path is divided into 20 phases and 511 lessons, each with a concrete deliverable. How it works: Each lesson follows a strict loop: read the problem, derive the math, write the code, run the provided test suite, and keep the resulting code artifact. Progress is tracked, and phases are marked as complete only when all constituent lessons are finished.
5. Extended Tracks for Specialization & Certification Prep: Beyond the core curriculum, it offers focused paths for specialization. How it works: Paths include "Model Context Protocol (MCP)" for building tool-integration systems, "Agent Skills" for creating portable agent capabilities, and practical labs for preparing for industry certifications (e.g., Claude certifications) with original practice questions and hands-on projects.

Problems Solved

1. Pain Point: It addresses the "tutorial hopping" problem and superficial understanding in AI education, where learners can deploy a model using libraries but cannot debug its loss curve, explain internal mechanisms like attention, or build systems without heavy reliance on abstractions.
2. Target Audience: Aspiring and current AI Engineers, Machine Learning Engineers, Software Developers transitioning into AI, Computer Science students, and Researchers seeking a stronger implementation foundation. It is specifically designed for individuals who prefer hands-on, code-first learning over passive video consumption.
3. Use Cases: A software engineer needing to understand transformer architectures deeply to optimize model inference. A student building a portfolio of fundamental AI implementations from scratch. A professional preparing for a technical machine learning engineering interview requiring low-level knowledge. A team lead establishing a standardized, foundational training resource for their engineering team.

Unique Advantages

1. Differentiation: Unlike most online courses, MOOCs, or documentation that start with teaching a framework (e.g., "Intro to PyTorch"), this curriculum intentionally delays framework use. It differs from academic textbooks by being entirely practical, code-driven, and hosted on GitHub with runnable tests. It is more structured and comprehensive than isolated blog posts or paper implementations.
2. Key Innovation: The integration of the "Skill" system for AI-native, interactive tutoring directly in the developer's terminal and IDE. This combines the depth of a textbook, the structure of a curriculum, and the interactivity of a personal tutor, all within an open-source, executable repository. The multi-language approach from the outset is also a unique practical feature for engineering-focused learning.

Frequently Asked Questions (FAQ)

1. Is AI Engineering from Scratch suitable for complete beginners? It is designed for learners with some programming experience (in one of the four languages) and high-school level mathematics. While it starts from the basics, its pace and depth are aimed at engineers and motivated students; true beginners may need to supplement with introductory programming or math resources first.
2. How does this curriculum compare to fast.ai or Andrew Ng's Deep Learning Specialization? While fast.ai is highly practical and top-down (starting with high-level APIs), and Andrew Ng's courses are theoretical and bottom-up, AI Engineering from Scratch is intensely implementation-focused and bottom-up. You spend most of your time writing the actual algorithms, not just using them or understanding their theory. It complements these resources by filling the "implementation gap."
3. Do I need to know all four programming languages (Python, TypeScript, Rust, Julia)? No. The curriculum is designed so you can follow the core concepts primarily through Python, which is the primary instructional language. The other language implementations are provided for additional perspective, performance comparison, or for learners who work in those ecosystems and want to see the concepts applied there.
4. Is this really free and open source? Can I use it commercially? Yes. The entire curriculum, including all lesson text, code, tests, and figures, is open-source under the MIT License on GitHub. This means you can use, copy, modify, and even use it for commercial purposes freely, forever, without any paywall or required sign-up.
5. What is the time commitment to complete the entire AI Engineering from Scratch curriculum? The curriculum is massive (20 phases, 511 lessons) and is designed as a deep, foundational journey, likely taking several months of dedicated part-time study or longer to complete fully. However, it is modular, allowing learners to target specific phases (e.g., "Ship model features" for LLM engineering) based on their goals.