Insights
Notes on AI orchestration, retrieval, evaluation, and what it takes to run systems in production.
-
AI orchestrationParallel agents, isolated worktrees: where AI orchestration actually buys you speed
Every coding tool shipped parallel agents and git-worktree support in 2026. The speed isn't in running more agents — it's in isolation, durable orchestration, and reviewing outcomes instead of process. The mechanics, and why I built IW AI Core around them.
-
BenchmarksI benchmarked my own RAG system on one GPU — here's what the numbers actually said
A small, honest measurement run on IW RAG: a retrieval ablation, eight local models compared, a quantization ladder, and a GraphRAG mode that scored zero. The reasoning matters more than the digits — and the sample is deliberately tiny.
-
RetrievalI built a local RAG system 14 months ago — here's how I'd build it today
RAG matured, and AI agents changed how you build. What I'd change about CORE, what I'd keep, and the part agents didn't speed up: the judgment about where a system must not be trusted.
-
Case studiesBuilding CORE: trusting LLMs where being wrong costs money
I built a 100%-local, five-agent AI platform for a regulated enterprise. The lasting lesson wasn't the demo — it was knowing when to trust an LLM where errors cost real money, and when not to.
Nothing here yet — try another topic.

Building something that has to be right?
Selectively available for engagements where owning the architecture of a production AI or BSS system is the point.
Start a conversation