The LLM Course That Treats LangChain as a Crutch

What it does

This repository holds the open-sourced code and slides from Hamza Farooq’s course on building LLM-powered systems by hand. It spans six weeks of material—tokenization, transformers, semantic search, custom RAG pipelines, fine-tuning, and local deployment—aimed at developers who already know Python and basic machine learning. The materials have been used in sessions at Stanford, UCLA, and MAVEN.

The interesting bit

The curriculum deliberately avoids LangChain and LlamaIndex, treating them as abstractions that obscure the underlying retrieval and generation mechanics. That makes it a rarity: coursework that forces you to write the plumbing yourself instead of importing it.

Key highlights

• 29 lessons and 6 projects covering transformers, vector search, and RAG without relying on pre-built frameworks
• Taught to over 1,500 professionals across Stanford, UCLA, University of Minnesota, and MAVEN
• Targets ML engineers and data scientists, not beginners; assumes Python and basic ML fluency
• Includes quantization, prompt-tuning, and deployment strategies for local and open-source models
• Syllabus culminates in building a custom search engine and a final demo project

Caveats

• The README advertises paid-course perks like live sessions, instructor access, and certificates alongside the “open sourced” materials, so it is unclear how much of the 29-lesson catalog is actually in the repository
• The GitHub description simply lists the repository as “Code and Slides,” which may not include all projects or guided content mentioned in the syllabus

Verdict

Worth bookmarking if you want to understand retrieval architecture and transformer internals without framework magic, but look elsewhere if you need a drop-in library or a beginner-friendly intro.