Handbook logo

Handbook

Quick tech ticks and comparisons

Generative AI

Learn about LLMs, LangChain, RAG, agents, and building AI-powered applications — explained in simple terms for everyone.

Recommended Learning Path:

  1. Part 1 - LangChain Foundations: Understanding LLMs and Orchestration — What an LLM is, why orchestration frameworks exist, and the core LangChain building blocks.
  2. Part 2 - LangChain Essentials: Building Blocks of AI Applications — Prompts, chains, and output parsers used in real applications.
  3. Part 3 - Tool Calling: Giving Your AI Real-World Superpowers — Letting an LLM call functions and APIs instead of just generating text.
  4. Part 4 - RAG Basics: Teaching AI About Your Documents — Retrieval-augmented generation: embeddings, vector stores, and grounding answers in your own data.
  5. Part 5 - LangGraph & Agents: Building Intelligent Workflows — Multi-step, stateful AI workflows and agent loops with LangGraph.
  6. Part 6 - Project: Build Your Own AI Agent — A hands-on project putting the previous five parts together.
  7. Part 7 - Observability with LangSmith: Debug and Monitor Your AI — Tracing, debugging, and monitoring LLM applications in production.

Deep Learning for Generative AI — Context Behind Transformers:

  1. Word Embeddings — How computers turn words into meaningful numbers: one-hot encoding, Word2Vec, cosine similarity, and the embedding layer.
  2. Encoder-Decoder Architecture — The seq2seq model: how an encoder compresses a sentence into a context vector and a decoder generates the output one word at a time.
  3. Attention-Based Encoder-Decoder — How attention lets the decoder look back at every encoder word instead of relying on a single vector.
  4. Math Behind Attention — Dot products, scaling, softmax, and the Query-Key-Value framework with worked numerical examples.
  5. Drawbacks of Attention-Based Encoder-Decoder — Sequential processing, vanishing gradients, and why the Transformer was invented.
  6. Introduction to the Transformer — Why encoder-decoder stays, what changes, and the group project analogy for Query, Key, and Value.
  7. Transformer Case Study — Python Program — A simple walk-through of the encoder on “Python program to print Hello World”, showing how each word attends to the others.
  8. The Transformer Architecture, Explained Simply — Every box in the Transformer diagram: embeddings, positional encoding, multi-head attention, add & norm, feed-forward, masking, cross-attention, and generating the output.
  9. The Math Behind Self-Attention — The magic is a matrix: X, W_Q, W_K, W_V and softmax(QKᵀ/√d)V, with a fully worked 3-word example.
  10. Self-Attention in Code with NumPy — Build self-attention from scratch in about 15 lines of NumPy on “I drink hot coffee”, with every output explained in plain words.
  11. Transformer Variants — Encoder-Only, Decoder-Only, Encoder-Decoder — Why the industry split the Transformer into three families: BERT for understanding, GPT for generating, T5 for transforming — and how to pick the right one.
  12. Sequence Classification with BERT (Hands-On) — Load a pre-trained BERT from Hugging Face, tokenize properly, fine-tune for sentiment classification with a plain PyTorch loop, and interpret the outputs.
  13. Sentence Embeddings with Sentence Transformers (Hands-On) — Turn whole sentences into vectors with all-mpnet-base-v2, then reuse the same embeddings for semantic search, document similarity, and k-means clustering — no training needed.
  14. Translation and Question Answering with Pre-Trained Models (Hands-On) — Use pre-trained MarianMT for machine translation and T5 for question answering, without training from scratch.

Multimodal Generative Architectures:

  1. Introduction to Multimodal Generative Architectures — Prediction vs transformation vs generation, discriminative vs generative models, conditional and unconditional generation, and why autoencoders reconstruct but cannot generate.
  2. Autoencoders and Variational Autoencoders — Why deterministic autoencoders fail at generation, the VAE probabilistic latent space, reparameterisation trick, ELBO objective, and a TensorFlow/Keras VAE on Fashion MNIST.
  3. Generative Adversarial Networks — Generator vs discriminator, the minimax game, conditional GANs, CycleGAN and cycle consistency.
  4. Diffusion Models and Stable Diffusion — Forward noising and reverse denoising, time conditioning, classifier-free guidance, and latent diffusion in Stable Diffusion.
  5. Multimodal Models and Vision Transformers — CLIP-style image-text alignment, prompts as steering vectors, cross-attention, Vision Transformers, and how modern text-to-image systems combine these pieces.
  6. Vision Transformers in Code (Hands-On) — Build a ViT from scratch in TensorFlow/Keras (patch extraction, encoding, transformer blocks), visualise image patches, and run Google’s pre-trained ViT from Hugging Face.

Short on time? Read the one-sitting recap instead (or first):

  • The Transformer - A Friendly Guide — The whole Transformer in one sitting with zero math: reading, writing one word at a time, training with cross-entropy loss, and what the “context limit” means inside each layer. Works as a standalone introduction or as a recap of Parts 6–11.

Opinion, case studies, and explainers:

Related Computer Vision topic:

  • Image Processing Fundamentals — Pixels, colour models, kernels, convolution, and feature extraction. Now the first entry in the Computer Vision series, since it’s foundational to vision rather than generative AI specifically.

What You’ll Learn:

  • What large language models are and how orchestration frameworks like LangChain fit around them
  • How to build prompts, chains, and structured output parsers
  • How to give an LLM tools and let it call real functions and APIs
  • How retrieval-augmented generation (RAG) grounds answers in your own documents
  • How to build multi-step, stateful agent workflows with LangGraph
  • How to trace, debug, and monitor LLM applications in production
  • How these ideas connect to real products and the broader direction of agentic AI