Building Agentic AI

Sinan Ozdemir is an AI expert and entrepreneur with a master's degree in pure mathematics from Johns Hopkins University. He founded Kylie.ai, patented agentic tool use there in 2018, participated in Y Combinator, and exited the company in 2019. Sinan is the author of Quick Start Guide to Large Language Models, Second Edition (Addison-Wesley, 2025), and cohosts the Practically Intelligent podcast. He has created several popular AI courses for Pearson on O'Reilly.

• Series Editor Foreword xiPreface xiiiAcknowledgments xviiAbout the Author xixPart I: Getting Started with Foundations of AI, LLMs, and Experimentation 1Chapter 1: An Introduction to AI, LLMs, and Agents 3Introduction 3The Basics of Large Language Models 3The Family Tree of LLM Tasks 10Alignment 10Prompt Engineering 12Special LLM Features 17LLM Workflows 25AI Agents 25Conclusion 28Chapter 2: First Steps with LLM Workflows 31Introduction 31Case Study 1: Text-to-SQL Workflow 32Conclusion 57Chapter 3: AI Evaluation Plus Experimentation 59Introduction 59Evaluating and Experimenting with LLMs 59Case Study 1, Revisited: The Text-to-SQL Workflow 61Case Study 2: A "Simple" Summary Prompt 77Conclusion 83Part II: Moving the Needle with AI Agents, Workflows, and Multimodality 85Chapter 4: First Steps with AI Agents and Multi-Agent Workloads 87Introduction 87Case Study 3: From RAG to Agents 88When Should You Use Workflows Versus Agents? 104Case Study 4: A (Nearly) End-to-End SDR 105Evaluating Agents 118Conclusion 121Chapter 5: Enhancing Agents with Prompting, Workflows, and More Agents 123Introduction 123Case Study 5: Agents Complying with Policies Plus Synthetic Data Generation 124Building Our Policy Bot Agent 127Case Study 6: Deep Research Plus Content Generation Agentic Workflows 133Multi-Agent Architectures 141Case Study 4, Revisited: Adding a Supervisor Agent to Our SDR Team 148Case Study 7: Agentic Tool Selection Performance 149Conclusion 157Chapter 6: Moving Beyond Natural Language: Multimodal and Coding AI 159Introduction 159Introduction to Multimodal AI 159Case Study 8: Image Retrieval Pipelines 168Case Study 9: Visual Q/A with Moondream 174Case Study 10: Coding Agent with Image Generation, File Use, and Moondream 176The Case for Any-to-Any Models 188Conclusion 191Part III: Optimizing Workloads with Fine-Tuning, Frameworks, and Reasoning LLMs 193Chapter 7: Reasoning LLMs and Computer Use 195Introduction 195Seven Pillars of Intelligence 195Case Study 11: Benchmarking Reasoning Models 198Reasoning Models for ReAct Agents 210Case Study 12: Computer Use 212Conclusion 224Chapter 8: Fine-Tuning AI for Calibrated Performance 225Introduction 225Case Study 13: Classification Versus Multiple Choice 227Case Study 14: Domain Adaptation 245Conclusion 258Chapter 9: Optimizing AI Models for Production 261Introduction 261Model Compression 261Case Study 15: Speculative Decoding with Qwen 269Case Study 16: Voice Bot--Need for Speed 272Case Study 17: Fine-Tuning Matryoshka Embeddings 277Case Study N + 1: What Comes Next? 284Index 287