Infrastructure and reliability engineering. Most of my work has been the unglamorous side of keeping things available: multi-region failover, secrets management, disaster recovery testing, and the evidence that proves any of it actually works.
I'm keenly interested in agent infrastructure, from the reliability and security side rather than the model side.
PipeRoll: a public registry of verified AI-agent incidents. Versioned schema, source-verified records, published corrections, CVE-style permanent identifiers. Records are CC BY 4.0; contributions welcome at piperoll/registry.
moat: sandboxed environment for coding agents. Zero-egress network behind a fail-closed proxy, credentials held outside the container, Terraform plan-only and kubectl read-only enforced host-side.
murmur: message bus for multi-agent sessions. Postgres-backed, long polling, single binary.
plumb: checks structured model output against the request that produced it, and repairs it in code instead of re-asking. Go, stdlib only.
The through-line: systems that act on their own should fail in ways you can see, bound, and recover from.