I build with Claude.
Then I ship it.

Problem. Companies run several AI agents, and none of them can prove its work was right.

Built. Self-hosted, MIT-licensed system that scores every agent against real team behaviour, promotes it through a shadow, propose, then act trust ladder, and keeps a verified-work ledger with EU AI Act Article 12 and 14 evidence.

Problem. Automation tools run steps, but nothing confirms the outcome you wanted was achieved.

Built. Describe an outcome, get a typed plan, rehearse it, approve it, run it, then verify it. A TypeScript monorepo with a Next.js app, Hono API, Better Auth, a worker and a CLI.

Problem. People who need sponsorship can't tell which employers actually sponsor, and job boards rarely say.

Built. Joined sponsor registers from the US (DOL, USCIS), UK, Canada, Ireland, Netherlands, Denmark and Germany to live listings from 675 employer job boards, refreshed hourly, with sponsor grading, in-browser CV matching, AI cover letters and a browser extension.

Problem. Citizen-science drug discovery games either fake the chemistry or need expensive GPUs.

Built. AutoDock Vina scoring and Monte Carlo pose search in plain JavaScript, four real disease targets, covalent warheads, a guided mode, a shared leaderboard and an AI lab partner served through its own Worker.

Problem. Over-the-counter tax-deed lists omit problems that make a parcel worthless.

Built. Eight state and federal sources in one subscriber product, with FEMA flood-zone and Census road-access checks, adjoining-parcel groups, an outcome ledger, a deal pipeline and a daily digest. Site and API run on one Cloudflare Worker; CI deploys and asserts the paywall.