Overview
AI engineering is the discipline of building AI systems from first principles — deriving algorithms from mathematics before introducing any framework. This contrasts with the prevalent tutorial-first approach of learning ML through frameworks (PyTorch, TensorFlow) without understanding underlying foundations.
The field spans from linear algebra foundations through neural network architectures, tokenization, attention mechanisms, and ultimately autonomous agent systems.
Math-first methodology
The core pedagogical principle: every algorithm is derived from raw math before a single framework is imported. Each lesson follows the pattern:
- Read the problem
- Derive the math
- Write the code
- Run the test
- Keep the artifact (runnable, multi-language implementation)
This approach is contrasted against fragmented AI education (scattered tutorials, framework-first shortcuts) to produce engineers who understand why algorithms work, not just how to call them.
Core topics
- Foundations: linear algebra, calculus, probability
- Core ML: backpropagation, gradient descent, supervised/unsupervised learning
- Deep learning: neural networks, tokenizers, attention mechanisms, transformers
- Applied: Python, TypeScript, Rust, Julia implementations
- Advanced: autonomous agent loops, swarm architectures
Open source curricula
AI Engineering from Scratch
- Maintainer: Rohit Ghumare (github.com/rohitg00)
- Scale: 503 lessons across 20 phases, MIT licensed, free
- Scope: “Linear algebra at one end, autonomous swarms at the other”
- Access: github.com/rohitg00/ai-engineering-from-scratch; runs locally, no paywall
- Languages: Python, TypeScript, Rust, Julia
Relationship to AI tooling
AI engineering is a prerequisite for understanding and effectively using higher-level tools like AI assistants, LLM wikis, and Claude Code. Understanding backpropagation and attention mechanisms from scratch enables informed decisions about model selection, fine-tuning, and system architecture.
Resources
- 2026-06-04 ◦ AI Engineering from Scratch — 503-lesson open-source curriculum, math-first approach, 20 phases from linear algebra to autonomous swarms, MIT license, Python/TypeScript/Rust/Julia