Lesson 44: State of the art in generative AI (LLM landscape and foundations)

Focus: Survey of state-of-the-art LLMs with foundational concepts for understanding and comparing models

Topic Purpose
The landscape Open vs closed weights, licensing, text and code generation models
Evaluating models Model sizes and parameters, context length, benchmarks
Looking ahead Key trends (efficiency, reasoning, agents) and limitations

Key concepts: Open vs closed weights, licensing, model sizes, context length, benchmarks, LLM landscape

Lesson 45: Practical LLM deployment (Working with large language models)

Focus: How to configure, interact with, and deploy LLMs - structured as a walk through the application stack

Topic Purpose
Model layer Decoding strategies, hyperparameters, performance optimizations
Inference layer Local (direct loading, Ollama), cloud, and API deployment options; tokens and cost
Framework layer Why use a framework; LangChain, LlamaIndex, Haystack, and others
Application layer System prompts, message format (system/user/assistant), streaming vs batch

Builds on 44: Applies foundation model knowledge to practical configuration and deployment

Key concepts: Application stack layers, decoding strategies, hyperparameters, deployment strategies, LLM frameworks, chat message format

Lesson 46: Advanced prompt engineering techniques: Part 1 (Prompting strategies)

Focus: Fundamental prompting approaches for eliciting better LLM responses

Topic Purpose
Zero-shot prompting Task completion without examples
Few-shot prompting Learning from in-context examples
Prompt structure Role definitions, temperature settings
Practical examples Translation, classification, text analysis

Builds on 45: Uses LLM APIs learned previously, focuses on prompt design

Key concepts: In-context learning, example-based prompting, prompt engineering fundamentals

Lesson 47: Advanced prompt engineering techniques: Part 2 (Advanced reasoning)

Focus: Sophisticated prompting techniques for complex reasoning tasks

Topic Purpose
Chain of thought (CoT) Step-by-step reasoning prompts
Self-consistency Multiple reasoning paths, majority voting
Tree of thoughts (ToT) Branching reasoning, exploration of solution space
Template formats Jinja2 templates, f-strings, custom templates
Dynamic message generation Programmatic prompt construction

Builds on 46: Extends basic prompting to complex multi-step reasoning

Key concepts: Reasoning chains, multi-path exploration, dynamic prompt generation

Lesson 48: LangChain for LLM application development: Part 1 (Core components)

Focus: Building blocks of LangChain applications

Topic Purpose
Models Chat models, LLM wrappers, direct API calls
Prompts ChatPromptTemplate, structured prompts
Output parsers Structured output extraction, format control
Basic chains Connecting models, prompts, and parsers

Builds on 47: Formalizes prompt engineering with LangChain abstractions

Key concepts: Model-Prompt-Parser pipeline, LangChain component architecture

Lesson 49: LangChain for LLM application development: Part 2 (Advanced patterns)

Focus: Retrieval, memory, and autonomous agents

Topic Purpose
Document loaders TextLoader, PyPDFLoader for various formats
Text splitters RecursiveCharacterTextSplitter for chunking
Embeddings HuggingFace and OpenAI embeddings
Vector stores Chroma for similarity search and persistence
Sequential chains Multi-step workflows with context passing
Memory Conversation history, state management
Agents Autonomous tool use (llm-math, wikipedia)
Local LLMs Running Falcon models locally

Builds on 48: Extends basic chains to RAG, memory, and agent patterns

Key concepts: Retrieval-augmented generation (RAG), vector databases, autonomous agents

Lesson 50: LLM fine-tuning, customization, and tool integration (Model adaptation)

Focus: Adapting pre-trained models and extending capabilities with tools

Topic Purpose
Data preparation Dataset formatting, preprocessing for fine-tuning
Fine-tuning techniques LoRA, QLoRA for parameter-efficient fine-tuning
Domain adaptation Adapting models to specific domains (legal, medical, etc.)
Evaluation Assessing fine-tuned model performance
Model Context Protocol (MCP) Standardized protocol for LLM-tool communication
MCP servers Building and deploying MCP servers for custom tools
Tool integration Connecting LLMs to external APIs, databases, and services
Function calling Structured tool use with modern LLMs

Builds on 49: Extends agent capabilities with custom tools and domain-specific models

Key concepts: Transfer learning, parameter-efficient fine-tuning, tool integration protocols, MCP architecture

Lesson 51: Advanced agents and production deployment (Capstone)

Focus: Complex multi-agent systems and deploying LLM applications to production

Topic Purpose
Multi-agent architectures Coordinating multiple specialized agents
Agent planning ReAct, reflection, and iterative refinement patterns
Agent memory systems Long-term memory, knowledge graphs for agents
Evaluation frameworks Testing agent reliability and performance
Production deployment Containerization, API design, monitoring
Cost optimization Caching, prompt optimization, model selection
Safety and alignment Guardrails, content filtering, output validation
Real-world applications Case studies of production agent systems

Builds on 50: Advanced patterns building on basic agents from Lesson 49

Key concepts: Multi-agent coordination, production engineering, safety considerations, deployment best practices

Course progression summary

Foundation (Lesson 44): LLM landscape, model sizes, context length, quantization, benchmarks
↓
Configuration & Deployment (Lesson 45): Decoding strategies, hyperparameters, chat vs completion, hosting options
↓
Prompt Engineering (Lessons 46-47): Zero-shot → Few-shot → CoT → ToT
↓
LangChain Development (Lessons 48-49): Core components → RAG + Agents
↓
Customization & Tools (Lesson 50): Fine-tuning, MCP servers, tool integration
↓
Advanced Production (Lesson 51): Multi-agent systems, deployment, safety

Key pedagogical flow: The course moves from surveying the generative AI landscape to practical LLM deployment, systematically building skills in prompting, application development, model customization, tool integration, and production deployment with advanced agent systems.