Writing
Notes on building these systems.
Mostly about what breaks in production and why the fix is usually structural rather than a better prompt.
The contract is the architecture
The most common mistake in production LLM systems: teams iterating on prompts when they should be iterating on output schemas.
Architecting multi-agent systems in production
~40 specialized LLM agents at Ember AI, orchestrated through LangGraph, with cost engineering that cut token spend by ~60%.
I shipped a public RAG demo
What's in it, and why the failing eval cases stayed in the report instead of being tuned away.
I stopped writing code first
Eight months ago I changed how I work: specs as the unit of work, not keystrokes. What that shifted wasn't speed.
Cost engineering for LLM systems
The most underrated production skill in AI right now — and the levers that actually move token spend.
Four ways your RAG system silently fails in production
Every team I've seen ship RAG has been bitten by at least one of these. None of them throw an error.
Eight months without hand-writing production code
Still shipping more, faster, and at higher quality. Why 'AI writes the code' is the least interesting part of the workflow.