LLM Primer V Building Real-World LLM Applications: Designing, Evaluating, and Operating LLM Systems in Production

The model is probabilistic. The contract you ship to users is not. Production AI engineering is the discipline of closing that gap. LLM Primer V is the practitioner's playbook for taking an LLM from impressive demo to dependable production system. It walks the full stack of decisions an engineering team makes between "this works on my laptop" and "this serves a million requests a day without paging anyone at 3 a.m." — model selection and prompt engineering, retrieval and agents, evaluation and observability, security and cost optimization. It is the book you reach for when the demo works and you have to ship the thing. What's inside The five pillars of AI engineering: reliability, validation, caching, cost control, feedback loops Sampling strategies (temperature, top-k, top-p, min-p) and defensive prompt engineering Structured outputs with JSON mode, Pydantic, Zod, and Outlines — real code, copy-pasteable RAG pipeline construction: chunking strategies, query transformations, multimodal and SQL retrieval Agent architectures (ReAct), tool calling with idempotency keys, and three memory systems The RAG Evaluation Triad and LLM-as-a-judge — including where it breaks Observability stacks: LangSmith, Langfuse, Arize Phoenix, Helicone — when each fits Prompt injection mechanics, the OWASP LLM Top 10, and the 4-Layer Mitigation Matrix Semantic caching, token-based rate limiting, dynamic model routing Inference acceleration: KV cache, continuous batching, speculative decoding Includes five reference appendices: production and security checklists, framework and observability decision matrices, streaming protocols and structured-output patterns, rate-limiting and cost-management architecture, and a glossary of AI engineering metrics. Code examples in Python, TypeScript, JSON, and YAML — written to be runnable, not pseudo-code. Three hundred pages of real engineering. Written for engineers, technical product managers, and architects building LLM systems for production. Assumes general LLM familiarity at the level of Volume I. Volume V of the LLM Primer series.