/kb — notes I prepared

Deep Learning
to Agentic AI.

A from-scratch, mathematically rigorous course in note form. Every module derives the math step-by-step, shows a tiny worked numeric example, and gives runnable code. Intuition → Math → Example → Code → Pitfalls.

Start from the top Back home

// curriculum12 modules

Deep Learning Foundations

Neurons, activations, the forward pass, loss functions, backprop (full chain-rule derivation), optimizers SGD→Adam, init, regularization, batch/layer norm.

Convolutional Neural Networks

Convolution math, kernels, stride/padding, pooling, receptive fields, backprop through conv, classic architectures (LeNet→ResNet).

RNNs & LSTMs

Sequence modeling, BPTT, vanishing/exploding gradients, LSTM & GRU gate math, worked example.

Transformers

Self-attention from first principles, scaled dot-product, multi-head, positional encodings, the full block and its training math.

Encoder / Decoder Architectures

Encoder-only (BERT), decoder-only (GPT), encoder-decoder (T5), causal masking, and why each shape exists.

Retrieval-Augmented Generation

Embeddings, vector search math, chunking, retrievers, rerankers, and an end-to-end RAG pipeline.

LangChain

LCEL, runnables, prompts, chains, memory, tools — with full runnable code.

LangGraph

State graphs, nodes, edges, conditional routing, cycles, and agent loops.

Agentic AI

ReAct, planning, tool use, multi-agent systems, memory, and evaluation.

Math Appendix

Linear algebra, calculus, and probability refreshers used everywhere above.

Exercises & Solutions

Pencil-and-paper + coding problems with fully worked solutions, per module.

Advanced Deep Dives

FlashAttention (online softmax), RoPE (full derivation), a trainable LSTM from scratch with BPTT, and the KV-cache.

// suggested study order

Do 01 until backprop feels mechanical — everything else is built on it.
02 + 03 teach the two classic inductive biases: spatial and sequential.
04 + 05 are the heart of modern AI — spend the most time here.
06–09 are the applied layer that turns models into products.
Dip into 10 whenever a derivation uses math you want to re-ground.