Book
LLMOps
Managing Large Language Models in Production
Abi Aryan
Summary
Abi Aryan, founder of Abide AI and a machine learning engineer with close to a decade building production ML systems, wrote this for teams who got a prototype working and are now facing what happens once real users depend on it. Its starting claim is that traditional MLOps doesn't transfer: models hallucinate instead of simply being wrong, prompt injection is a new attack surface, and agents introduce failure modes that standard monitoring doesn't catch. From there it works through LLMOps team roles, the architecture of RAG and agent-based applications, and the data engineering underneath them, before turning to evaluation metrics, early detection of hallucination and other critical failures, governance and privacy, and building an observability pipeline that scales without the infrastructure bill scaling with it.
Target Readers
- ML engineers and SREs handed an LLM prototype and told to make it production-ready, with no playbook for what breaks differently from a traditional model
- Teams building RAG or agent-based applications who need an evaluation strategy for hallucination and failure, not just a demo that worked once
- Engineering leads defining governance and security review for LLM systems, including prompt injection, ahead of a production launch
Tags
Colophon
- Publisher
- O'Reilly Japan
- ISBN
- 978-4-8144-0160-4
- Published
- May 2026
- List price
- ¥5,280incl. taxMay differ from the actual selling price on Amazon
Get this book
* The link above is an advertisement via Amazon Associates.Related Books
Prerequisites
- Recommended
Observability Engineering
Achieving Production Excellence
Charity Majors, Liz Fong-Jones, George Miranda
Reason: Where 'Observability Engineering' teaches observability for distributed systems in general via structured events and distributed tracing, 'LLMOps' applies that thinking to running LLM-based applications, connecting it to LLM-specific observation challenges like evaluation-metric design and early detection of hallucination.
- Recommended
Seisei AI no Anzensei Nyumon (An Introduction to Generative AI Safety)
Koki Wataoka
Reason: Once you've learned generative-AI risk classification, red-teaming-based safety evaluation, and system-level countermeasures like guardrails, you advance to embedding them into a mechanism that actually keeps running in production. LLMOps' closing chapter covers responding to security risks such as governance, privacy, and prompt injection, plus building an observability pipeline — connecting the evaluation and countermeasure knowledge from the AI safety primer into concrete production-operation mechanisms.
- Related
Site Reliability Engineering
How Google Runs Production Systems
Betsy Beyer, Chris Jones, Jennifer Petoff, Niall Richard Murphy
Reason: Where Google's 'Site Reliability Engineering' frames engineering reliability for general services via SLIs/SLOs and error budgets, 'LLMOps' applies that operating philosophy to the new target of LLM applications, organizing the production concerns—including governance and cost management—specific to running them.