A practitioner's guide to building AI agent systems that actually work in production — closing the gap between demo quality and the reliability standards that financial services, healthcare, and regulated industries actually require.
Written from 16 years of production experience in high-stakes systems where being wrong has real consequences. Not a research paper. Not a tutorial. A field guide for engineers who have to ship and support the thing.
One email when the book is ready. No spam, no newsletter.
Why agents that work in demos fail in production — and the architectural patterns that close that gap without making the system fragile.
Financial services and healthcare have compliance requirements that most agent architectures ignore. How to design around them from the start, not retrofit later.
The real engineering challenges of multi-step agents — tool design, failure handling, state management, and avoiding the cascading failure modes most tutorials skip.
How to build eval harnesses for agents that produce meaningful signal — not just "did it answer" but "did it do the right thing, consistently, under load."
When and how to put humans back in the loop — without making the agent useless or the human a bottleneck.
The architecture decisions that need to happen before you go live — observability, versioning, rollback, cost control. The things you wish you'd thought about earlier.
You know how to build reliable systems. This book connects what you already know to what's different about agents — without making you feel like a beginner or pretending the hard parts don't exist.
You're the person who has to own the decision about how agents fit into an existing system. This book gives you the vocabulary and the tradeoffs, not just the happy path.
If your company operates under PCI-DSS, HIPAA, or SOX, most agent content online doesn't apply to you. This book starts with those constraints, not as an afterthought.
You need to make a call on where and how to use agents — and you need to be confident in that call. This book gives you the mental model to evaluate options, not just a vendor pitch.
Technical articles at DZone, InfoQ, and others covering event-driven systems, cloud-native patterns, and applied AI. Some in editorial review, more coming.
See all writing →Three production-quality tools including an AI Content Curator and the Drift Explainer with a real LLM eval harness — working examples of the patterns the book covers.
See all projects →