Field notes

What actually happens when you take GenAI to production — RAG that has to cite its sources, agents that have to stay deterministic, the AWS bill underneath. Written from systems that are live, not from a tutorial.

Putting AI into a legacy system without it falling over10 June 2026 · 5 min read

The hardest part of production GenAI usually isn’t the AI — it’s the ten-year-old stack it has to live inside. How I build a clean boundary the model can trust.

LEGACYARCHITECTUREINTEGRATION

Production RAG — what the tutorials don't show you8 June 2026 · 7 min read

Demos retrieve. Production cites, verifies, and refuses to be confidently wrong. The gap is everything I learned shipping Preamble.

RAGPRODUCTIONEVALS

How I evaluate GenAI — the metrics that actually matter5 June 2026 · 5 min read

Faithfulness, context precision/recall, answer relevancy, and a golden set you watch on every change. How "it seemed fine" becomes "it’s measurably good."

EVALSRAGPRODUCTION

Bedrock or the Anthropic API directly — when each wins2 June 2026 · 5 min read

It's not a model decision, it's a constraints decision. A field guide from shipping both.

ARCHITECTUREAWSLLM

The six pillars of production GenAI, in practice28 May 2026 · 5 min read

The AWS Well-Architected GenAI Lens names the unglamorous things that separate a demo from a system you can put in front of customers. What each pillar means on a real build.

AWSPRODUCTIONARCHITECTURE

RAG or fine-tuning — how I actually decide20 May 2026 · 4 min read

Fine-tuning teaches a model how to behave; RAG gives it what to know. Most problems are knowledge problems. Here’s the order I go in.

RAGFINE-TUNINGARCHITECTURE

// get the next one

One email when a new piece drops. No spam, no list-churn.