Courses › Course06

Course06 · LLMs and Generative AI

Large language models as a working material: how they are made, how to prompt and call them, how to ground them with retrieval, customize them, and go multimodal.

49 days 49 written 49 complete

Days 239-245 · The LLM Landscape 7/7 complete
  1. Day 239 How Large Language Models Are Trained complete
  2. Day 240 Pretraining, Fine-Tuning, and RLHF complete
  3. Day 241 The Model Landscape: Claude, GPT, Gemini, Llama complete
  4. Day 242 Open Weights versus Closed APIs complete
  5. Day 243 Tokens, Context Windows, and Sampling complete
  6. Day 244 Capabilities, Limits, and Hallucination complete
  7. Day 245 Benchmarking Models Yourself complete

Project: Model Comparison Study — Design a ten-task benchmark and compare two accessible models on it, documenting failure patterns and cost.

Days 246-252 · Prompt Engineering 7/7 complete
  1. Day 246 Prompting Fundamentals complete
  2. Day 247 System Prompts and Role Design complete
  3. Day 248 Few-Shot Examples and Chain of Thought complete
  4. Day 249 Structured Output: Getting Reliable JSON complete
  5. Day 250 Prompt Patterns and Templates complete
  6. Day 251 Prompt Injection and Safe Prompting complete
  7. Day 252 A Tested Prompt Library complete

Project: Reusable Prompt Library — Build a tested library of parameterized prompts for five recurring tasks, each with evaluation examples.

Days 253-259 · LLM APIs 7/7 complete
  1. Day 253 First Calls to the Claude API complete
  2. Day 254 The OpenAI-Compatible Ecosystem complete
  3. Day 255 Streaming Responses complete
  4. Day 256 Tool Use and Function Calling complete
  5. Day 257 Working with Images and Documents complete
  6. Day 258 Cost, Caching, and Rate Limits complete
  7. Day 259 Building a CLI Assistant complete

Project: Command-Line AI Assistant — Build a streaming CLI assistant with conversation memory, tool use, and a cost meter.

Days 260-266 · Embeddings and Vector Search 7/7 complete
  1. Day 260 What Embeddings Are complete
  2. Day 261 Semantic Similarity Search complete
  3. Day 262 Vector Databases complete
  4. Day 263 Chunking Strategies complete
  5. Day 264 Hybrid Search and Rerankers complete
  6. Day 265 Evaluating Retrieval Quality complete
  7. Day 266 Semantic Search over Your Own Notes complete

Project: Semantic Search Engine — Build semantic search over your own notes with embeddings, a vector store, and a retrieval-quality evaluation.

Days 267-273 · Retrieval-Augmented Generation 7/7 complete
  1. Day 267 The RAG Architecture complete
  2. Day 268 A Minimal RAG System from Scratch complete
  3. Day 269 RAG over PDFs and Messy Documents complete
  4. Day 270 Citations and Grounded Answers complete
  5. Day 271 Advanced RAG Patterns complete
  6. Day 272 Evaluating RAG Systems complete
  7. Day 273 A Documentation Assistant complete

Project: Documentation Q&A Bot — Build a RAG assistant over a real documentation set with citations and a measured answer-quality eval.

Days 274-280 · Customizing and Running Models 7/7 complete
  1. Day 274 Prompting versus RAG versus Fine-Tuning complete
  2. Day 275 Fine-Tuning with LoRA complete
  3. Day 276 Building Fine-Tuning Datasets complete
  4. Day 277 Running Local Models with Ollama complete
  5. Day 278 Quantized Inference and llama.cpp complete
  6. Day 279 Serving Open Models complete
  7. Day 280 Fine-Tune and Serve Your Own Model complete

Project: Local Fine-Tuned Model — Fine-tune a small open model with LoRA on a task dataset and serve it locally, comparing before and after.

Days 281-287 · Multimodal and Generative Media 7/7 complete
  1. Day 281 How Diffusion Models Generate Images complete
  2. Day 282 Image Generation in Practice complete
  3. Day 283 Speech: Recognition and Synthesis complete
  4. Day 284 Video and Music Generation complete
  5. Day 285 Multimodal Models complete
  6. Day 286 Generative AI Ethics and Copyright complete
  7. Day 287 Section Project: A Multimodal Application complete

Project: Section Project: Multimodal App — Build an application combining at least two modalities — for example, speech in, grounded text out, generated image alongside.