I build systems that catch confident AI failures.

What I build

I build systems that catch confident AI failures. The work on this site spans enterprise AI delivery, a shipped iPhone product, and the agents I use every day. I take messy business problems and turn them into working products, semantic systems, and agent workflows. Approval, evidence, and safe failure are built into the way they operate.

Enterprise AI

In my day job, I build the delivery system behind enterprise AI. It starts with discovery and semantic modeling. From there, I build the Claude Code workflows, evaluation, approval gates, reversible writes, rollback, and final browser checks that make the system safe to operate.

Those controls caught two different ways a quality gate could approve the wrong answer. In one case, a query ran successfully but returned the wrong result. In another, a cached response made a failed fix look successful. I fixed both problems and added named regression tests so they stay fixed.

The same platform binds every change to the right target, keeps a person in the approval loop, makes writes reversible, and checks the result in the browser before users inherit the risk.

Ed and Taz

I built Ed on OpenClaw as a persistent personal agent. Ed keeps approved context, uses tools, handles scheduled work, and follows operating rules I defined.

My current agent environment runs on Hermes Agent, an open-source framework from Nous Research. I gave it durable memory and reusable skills. It can delegate coding work to local models, automate browsers and desktop apps, run scheduled tasks, and keep family and business profiles separate.

Taz is the private family coordination agent I built inside that environment. She works within a separate family boundary. She can use approved calendars and school information, create protected family files, and prepare messages. A parent must approve consequential actions before she takes them.

DimeVision

I built DimeVision, an AI-assisted coaching product for iPhone. A welder photographs a weld, and the app evaluates visible surface characteristics and returns practical feedback.

DimeVision is a training aid. It does not replace qualified inspection. I took it from industry discovery, through product development, to a working App Store release.

What I bring

My best work sits where business, data, product, and engineering meet. I take incomplete requirements, find the real problem behind them, and build a system that works in practice. Then I test it against real behavior and stay accountable through delivery.

I add the most value when a promising demonstration has to become something people can actually use and trust.