These are the tools I build in the open. They are not client work, but they are built to the same rule: a deterministic core that behaves the same on every run, a model only where judgment is needed, and the last call left to a person. Shipped client systems sit under case studies.
Running an AI pilot that is not production-ready yet? That is exactly what I do: audit, fixed-price scope, delivery in 2–6 weeks. That includes the part a pilot never shows: state that survives between runs and stays auditable, approvals where ownership is required, and a way back when a decision turns out wrong.
I use analytics (PostHog, EU-hosted) only with your consent to improve this site. Essential functions always work. See the privacy policy.