LLM Primer V — Series Introduction & Index

A chapter-by-chapter walkthrough of LLM Primer V: Building Real-World LLM Applications — the volume that treats AI engineering as a discipline in its own right, not a prompt trick.

Why this series exists

There is a recurring pattern in applied AI work: a team builds a demo in a weekend, the demo dazzles, a pilot is approved, and six months later the pilot is stuck in the phase that the team calls "hardening" and privately calls "the part where the model keeps doing things we did not expect." Nothing about the model changed. What changed is that the work moved out of the region where a handful of curated inputs produce plausible outputs, and into the region where thousands of real users, adversarial content, cost lines, and latency budgets set the terms. Volume V is about the discipline that lives in that second region. It is not a book about training models, and it is not a book of prompt tricks. It is a book about the engineering that turns a model into a system a business can operate.

The book in one sentence: AI engineering is the discipline of building deterministic systems around probabilistic models, and this volume walks the eight surfaces — foundation models, prompts, retrieval, agents, evals, observability, security, and serving — where that discipline lives.

Who I wrote this for

Application engineers who have shipped a working prototype and are now being asked to keep it up. Tech leads deciding whether a new LLM feature is worth the operational burden. Product managers on AI teams who want to understand what their engineers are trading off between latency, cost, and quality. The book assumes you have used an LLM API and read enough of a framework's docs to know what a prompt template is. It does not assume any particular background in ML research; the mechanics that matter for production are engineering mechanics, and the book stays in that register.

How to read it

The chapters are ordered as a stack. Chapter 1 sets the frame — the deterministic wrapper around the probabilistic core — and the rest of the book fills in the wrapper's layers. Chapters 2 through 4 walk the model, retrieval, and agents in turn; Chapters 5 and 6 add the evaluation and observability rails that let the system evolve safely; Chapters 7 and 8 close with security and serving economics. You can read the book straight through, or you can jump into the chapter that matches the failure mode your system is currently exhibiting. Each chapter's walkthrough article distills three key ideas and points at the source material.

The 8-chapter walk

The walkthrough runs one article per chapter, published daily.

April 14 — Chapter 1 — The Discipline of AI Engineering. The reliability gap between demo and production, and the five pillars that close it.
April 15 — Chapter 2 — Foundation Models & Prompt Engineering. Model tiering, sampling parameters, defensive prompts, and structured outputs as engineering surfaces.
April 16 — Chapter 3 — Retrieval-Augmented Generation. The RAG pipeline end to end: chunking, hybrid retrieval, query transformation, multimodal, and text-to-SQL.
April 17 — Chapter 4 — AI Agents and Tool Calling. ReAct loops, tool schemas as contracts, and the three memory layers agents actually need.
April 18 — Chapter 5 — Evaluating LLM Applications. LLM-as-judge, the RAG Triad, and trajectory tests for agents.
April 19 — Chapter 6 — AI Observability and Tracing. OpenTelemetry GenAI conventions, the metrics that matter, and the loop back into evals.
April 20 — Chapter 7 — LLM Security and Guardrails. The OWASP Top 10, direct-versus-indirect injection, and the four-layer mitigation matrix.
April 21 — Chapter 8 — Optimizing Performance, Serving, and Cost. Semantic caching, dynamic routing, and what actually happens inside the inference server.

Where this sits in the series: Volumes I and II covered the mechanics of the models themselves, Volume III went deep on retrieval, and Volume IV built the MCP-shaped cognitive architecture; Volume V is where those threads become a single production system, and Volume VI will take that system to fleet scale.

About this book and the series

The LLM Primer is a seven-volume series that treats language models as engineering rather than magic. Volume V is the fifth entry. The house style is sober and mechanism-first: every chapter names the failure mode it is closing, walks the mechanism that closes it, and shows the code where the mechanism lives. There are no exhortations to embrace AI. The reader is assumed to have already decided that; the book is about doing it well. Later volumes go deeper on scaling (VI) and on the harder security and governance questions that regulated deployments face (VII).

Grab a copy. The paperback includes the full appendices — deployment topologies, framework comparisons, cost worksheets — that the walkthroughs summarize but do not reproduce. LLM Primer V on Amazon →