Writing
Production forensics and engineering notes from building and shipping AI products.
Surviving a Model Deprecation: Splitting One Persona Across Three Models to Cut Costs
When our primary model was deprecated, upgrading meant a massive cost explosion. Here is how I built a safety gauntlet and orchestrated a three-model pipeline to cut costs and increase accuracy.
Designing for Distrust: Why AI Adoption Requires Verifiability Over Speed
Generating a fast answer is easy. Giving non-technical users the tools to verify that answer even faster than they could without AI is the real breakthrough for adoption.
Same payload, two backends: Vertex cached nothing, Studio cached every turn
My Gemini cache hit count was zero on my own setup, while the same prompt caches every turn on a sibling backend. The forensic, the paired test that proves it's the backend, and the fix.