AI product engineer. I take an agent idea from an empty repo to a live thing you can click — owning the product call, the UX, and the proof it works. USC Business + AI '24. Building and shipping AI products.
- Farthing — an AI support agent allowed to touch money: the model can't set the amount, a policy engine decides it, a human approves, and the Stripe refund can never execute twice → https://github.com/theo-ai-lab/farthing
- Toffoli — an undo layer for AI agents: flags which actions are reversible, compensable, or permanent, and escalates the rest to a human → https://theo-ai-lab.github.io/toffoli/
| Product | What you get | Try it |
|---|---|---|
| Farthing | A Stripe refund agent where the model can't set the amount — policy decides, a human approves, and no refund can execute twice or without approval (Stripe test mode) | repo |
| AI Career Coach | Resume-grounded coaching, no signup | live |
| Pacioli | Checks what your agent claimed against the evidence, prints a receipt | live |
| Toffoli | An undo button for AI agents — flags which actions are safely reversible | demo |
| Maudslay | Go / no-go before you ship a computer-use agent | repo |
| Plimsoll | Catch agent regressions in CI in seconds — before users see them | demo |
| Coehoorn | Red-team your agent before your users do | demo |
I ship fast and I prove it — each product ships with a real eval, red-team, or reconciliation, so the quality claim is falsifiable, not vibes. The AI proposes; the code disposes.