I direct AI coding agents (mostly Claude Code, sometimes Gemini) and decide what they build. Nearly everything on this account was written by AI under my direction, and I label it that way.
- the-penniless-agent: can an AI agent with no money and no accounts earn money for an ordinary person? A running field report that writes down every blocker in the way. So far: $0.00.
- boardcheck: checks AI-agent job-board listings before you bid, flagging fake engagement, unpaid work, upfront fees and prompt injection. A command-line tool and an MCP server listed in the official MCP Registry, with a browser demo.
- alba-duplex-research: a full-duplex voice model (it listens and talks at the same time) on a 4 GB laptop GPU, with code, tests and a review paper. Paused and unfinished; help is welcome.
- was-my-github-hacked: a plain 15-minute check for a hijacked GitHub account, and what to do, in order, if it was. Built on signs that last from one attack wave to the next.
Also: echo-earning-agent (a daily watcher that checks whether money or work has arrived) and ai-money-safeguards (what has to be settled before AI handles people's money and benefits).
Retired: token-intel-x402 (offline, kept as reference code) and boardcheck-edge (a prototype that never launched).
Finding where AI-assisted contributions actually help open-source maintainers, and keeping those contributions small, tested and clearly labeled. Recently merged: lhotse #1641 and #1607 (speech-data tools). The full record, including what didn't work, is in CONTRIBUTIONS.md.