All Writing
Everything I've written — deep dives, project breakdowns, and the thinking behind the work.
The Forge: Knowledge That Compounds
I got tired of re-teaching AI agents the same lessons every session. So I built a system where every correction, pattern, and hard-won fix compounds instead of evaporating.
The Prompt Is Not the Product
After 25 years of shipping products, I can tell you: the prompt is the easy part. The hard part is everything you build around it when it inevitably fails.
Evals Are Unit Tests for Vibes
You can't assert on 'sounds helpful but not condescending.' After thousands of evals, here's what actually works — and the trap that wastes everyone's time.
Context Windows Are a Lie (Sort Of)
128k tokens is a marketing number. The usable window is a fraction of that — and I learned the hard way that how you fill it matters more than how much you have.
Ship the Feature Without the AI First
Counterintuitive? Maybe. But after 25 years of shipping products, I know that the non-AI version teaches you everything the AI version needs to know.
The Temperature Misconception
Most developers treat temperature as a creativity dial. Twenty-five years of debugging has taught me to distrust any parameter I can't fully explain — and temperature is a perfect example.
RAG Is Not a Search Engine
Every tutorial makes RAG look like three steps. After building retrieval pipelines that actually work in production, I can tell you it's more like thirty.
Why Fraud Detection Must Ask Questions, Not Give Answers
I built a fraud detection system that replaces confidence scores with structured doubt. Here's why single-number risk scores are the most dangerous thing in compliance — and the architecture that fixes it.