Exploring AI safety and accountability through local-first systems. I build experimental tools for inspectable agent memory, explicit permissions, and verifiable execution records.
I’m interested in systems people can inspect: where a remembered claim came from, what a check actually established, and who authorised an action. My work explores these questions through small, testable implementations.
| Project | Focus |
|---|---|
| Magpie | A replayable memory kernel for AI systems. Signed history, explicit evidence, and policy-defined conclusions. |
| Deadbolt | A permission engine for AI-operated systems. Models propose actions; authority remains outside the model. |
| Cogitator | A Rust recorder and verifier for agent runs, with tamper-evident records and reproducible verification. |
These are experimental research projects, developed with an emphasis on clear boundaries, adversarial testing, and documented limitations.