Developing Advanced Generative AI Applications

Course Overview

Module 1 - Parameter-Efficient Finetuning Techniques

By the end of this module, you will be able to recognize when PEFT beats full finetuning, apply prompt-based and low-rank adaptation techniques, choose between LoRA, IA3, and prompt-tuning for a given task, and evaluate PEFT performance against the base model.

• Parameter-efficient finetuning vs. traditional finetuning
• Prompt-based methods of finetuning
• Low-rank adaptation techniques
• IA3 and other techniques
• Selecting a PEFT technique
• Evaluating PEFT performance
• Hands-on Lab: Apply prompt-tuning with HuggingFace PEFT to harden one model against indirect prompt injection.

Module 2 - Chatbots with Memory and Persistence

By the end of this module, you will be able to persist chatbot conversation state, manage long conversations through summarization, retain memory across multiple interactions, extract and verify facts, and handle privacy in chatbot memory.

• Memory and Persistence in Chatbots
• Persisting Messages within a Conversation
• Handling Long Conversations
• Memory Across Multiple Interactions
• Fact Extraction and Verification
• Memory and Privacy
• Hands-on Lab: Build a chatbot with persistent history that remembers facts and context across multiple sessions.

Module 3 - Multi-modal Prompting and Interaction

By the end of this module, you will be able to process non-textual inputs, prompt to generate and combine images and text, apply Stable Diffusion for image generation, and reduce multi-modal generation costs.

• How Non-Textual Inputs are Processed
• Prompting to Generate Images
• Prompting with Images and Text
• Understanding Stable Diffusion
• Reducing Costs of Multi-modal Generation
• Use Cases for Multi-modal Prompting
• Hands-on Lab: Build a multi-modal agent that improves accessibility through combined image and text generation.

Module 4 - Streaming, Ensembling, and Advanced Output Techniques

By the end of this module, you will be able to stream LLM outputs, structure pipeline progression for streaming, apply self-consistency and self-refinement, and use mixture methods for ensembled output.

• Streaming LLM Outputs
• Streaming Pipeline Progression
• Self-Consistency
• Post-Processing of Generated Content
• Mixture Methods
• Self-Refinement
• Hands-on Lab: Build a streaming LLM pipeline with self-consistency and self-refinement of mixed outputs.

Module 5 - Function and Tool Calling

By the end of this module, you will be able to bind functions to LLMs, handle tool-calling responses, apply error handling and retries, secure LLM function calls, and accelerate tool use with LangChain ToolKits.

• How does Function Calling Work?
• Binding Functions to LLMs
• Handling LLM Responses
• Error Handling and Retries
• Security of LLM Function Calls
• LangChain ToolKits
• Hands-on Lab: Build a function-calling agent that interacts with an external environment through LangChain ToolKits.

Module 6 - Advanced Generative AI Use Cases

By the end of this module, you will be able to apply LLMs to embedded agents, agent-based simulation, automated code generation, cybersecurity, and risk management — including the EU AI Act categories of risk.

• High-Efficiency Embedded Agents
• Agent-based Modeling and Simulation
• Automated Code Generation and Execution
• Cybersecurity Applications
• Risk Management Use Cases
• EU AI Act Categories of Risk
• Hands-on Lab: Build an LLM-based simulation and identify risks against the EU AI Act categories.

Module 7 - Prompt Chaining, Routing, and Directed Graphs

By the end of this module, you will be able to compose prompt chains with predetermined and generated paths, pass context between prompts, extend chains into directed graphs, apply conditional routing and cyclic graphs, and debug complex chains.

• Prompt Chaining with Predetermined Paths
• Passing Context between Prompts
• Prompt Chaining with Generated Paths
• Extending Prompt Chains with Directed Graphs
• Conditional Routing and Cyclic Graphs
• Debugging Complex Prompt Chains
• Hands-on Lab: Build a ReAct agent with LangGraph that solves the Zebra Problem with generated paths.

Module 8 - Introduction to Agentic AI and Autonomous Systems

By the end of this module, you will be able to define agentic AI, differentiate it from traditional AI and RPA, articulate its benefits and challenges, and reason about its impact on key sectors and complex problem solving.

• Defining Agentic AI and its Key Characteristics
• Differentiating Agentic AI from Traditional AI and Robotic Process Automation
• Understanding the benefits and challenges of Agentic AI
• Real-world applications of Agentic AI
• Potential Impact of Agentic AI on Key Sectors
• Agentic AI for Solving Complex Problems
• Hands-on Lab: Implement a simple agent in Python and exercise round-robin communication between two agents.

Module 9 - Agent Architectures and Frameworks

By the end of this module, you will be able to compare reactive, deliberative, and hybrid agent architectures, design a blend of reactive and deliberative behaviour, and choose between AutoGen, LangGraph, and CrewAI for a given application.

• Overview of Architectures - Reactive, Deliberative, and Hybrid
• Comparison of Architectures
• Exploring Reactive Architectures
• Exploring Deliberative Architectures
• Blending Reactive and Deliberative Architectures
• Agentic Frameworks - AutoGen, LangGraph, CrewAI, etc.
• Hands-on Lab: Implement a reactive agent in AutoGen alongside a deliberative agent in LangGraph and compare their behaviour.

Prerequisites

Participants should enter this course with:

Practical Python experience
GAI-1101 or equivalent
​GAI-2101 or equivalent