Part 3 Reliable operations
You’ve built something that works. Your LLM produces accurate outputs. Your agents take safe actions. You ship to production on a Friday and go home feeling great. Then Monday arrives. A model provider quietly updates its weights, and your intent classifier starts routing 30% of queries to the wrong agent. A user discovers that a carefully worded prompt bypasses your safety filters. Your costs triple because a feedback loop in one agent keeps calling the API in circles. None of these situations was a problem in staging.
The final part of this book is about what happens after deployment. Your users don’t care about your architecture; they care that the system works today the same way it worked yesterday and that it treats them fairly. This section gives you the tools to make that happen.
Chapter 9 builds your evaluation toolkit: LLM-as-a-judge; FActScore; trajectory analysis; and the streaming, batching, and caching patterns that make production systems fast and cost-effective. Chapter 10 covers LLMOps: deployment architectures, real-time monitoring, hybrid routing between cloud and self-hosted models, and knowing when to upgrade or switch models. Chapter 11 closes with the human side: bias detection, privacy protection, multilayered safety architectures, and the regulatory landscape, including the European Union’s AI Act and Colorado’s AI Act.