From Derivatives to DeepSeek: A Notebook Curriculum for AI Engineering

What it does This repository is a self-contained curriculum of Jupyter notebooks designed to take you from mathematical foundations through to working AI systems. You start with linear algebra and probability, move into PyTorch tensor manipulation, then build neurons, layers, and full neural networks from scratch—including modern touches like RMSNorm and the Muon optimizer. Later modules cover decoder-only transformers, end-to-end RAG pipelines with vector stores and embedding models, and OCR utilities for text extraction.

The interesting bit Most learning repos either stay in the shallow end with scikit-learn or throw you straight into fine-tuning scripts. This one attempts the full stack: you derive gradients before you ever call nn.Linear, then assemble multi-head attention by hand before pointing an LLM at a vector database. It is an ambitious scope for a single repo, essentially trying to replace a semester of coursework with well-organized .ipynb files.

Key highlights

• Progression from math fundamentals → PyTorch → neural networks → transformers → RAG → OCR
• Neural network module builds from scratch, covering activation functions, normalization, and modern optimizers (Adam, Muon)
• RAG module includes chunking strategies, retrieval, and integrations with cloud LLMs (Atlas Cloud, MiniMax, OpenAI-compatible APIs)
• Also includes traditional ML implementations: linear regression, logistic regression, decision trees, and naive Bayes
• Provides a recommended learning_path.md for step-by-step progression

Caveats

• Dependency management is fragmented: the root has a requirements.txt, but subfolders like 5.RAG/ and 6.OCR/ maintain their own.

Verdict Grab this if you are a beginner or mid-level engineer who wants a structured, hands-on tour from math to modern LLM infrastructure. Skip it if you are hunting for a single deployable tool or a framework to import—this is a course in repo form.