Blog

Notes taken while building.

Short pieces on approval queues, evaluating agents, and what open models actually cost to run.

August 2026

The approval queue is the product

Most AI automation projects fail on trust, not capability. The interface where a human reviews what the agent proposes is the part that decides whether anyone uses it.

infrastructure

July 2026

When self-hosting an open model actually pays

There is a token volume above which running your own model on vLLM is cheaper than an API, and a compliance threshold below which it is the only option. Here is how to work out both.