When an AI system fails in production, the model is rarely the problem. The system around it is.

AI Systems Design Patterns is a problem-first catalog for the engineers who have to make LLMs reliable, affordable, and trustworthy at scale — long after the demo has stopped impressing people.

Instead of organizing by framework or the model of the month, it organizes by the five forces that act on every AI system: non-determinism, cost, latency, trust, and drift. Name the force that's hurting, and turn straight to the patterns that move it.

Inside you'll find:

29 production patterns across retrieval and grounding, generation and control, orchestration, reliability and operations, and data and feedback — RAG, structured output, validate-and-repair, model routing, guardrails, semantic caching, drift detection, LLM-as-judge, and more.
12 anti-patterns given equal weight — Confident Hallucination, Retry Storms, Ground-Truth Theater, Over-Agenting, and the other failure modes named, diagnosed, and cured.
32 diagrams that carry the argument, not decorate it.
One continuous worked example — a document-intelligence platform threaded through every pattern, so you see how the pieces actually compose into a working system rather than a pile of disconnected tricks.
A Reference Shelf with a Pattern-Selection Map and a Trade-off Cheat-Sheet for fast decisions under pressure.

Every pattern is framework-agnostic pseudocode built to outlast this season's tooling, paired with honest trade-offs, a clear "when not to use it," and the production war stories behind it.

If you've shipped an AI feature and watched it wobble under real traffic — costs creeping, answers drifting, trust eroding — this is the field guide to building the system that holds.

For engineers, architects, and technical leads building production-grade AI and LLM systems. A companion to Agent Engineering: From Prototype to Production.