I build tools that make AI agents safer, and I write about what AI is doing to work.
Open to conversations about AI safety, security and applied ML roles.
Prompt-injection defense for LLM coding agents. A reference monitor tracks attacker-influenced data through a session and gates tool calls, with a content-blind LLM judge for the grey zone and a benchmark harness across 3,000+ attack scenarios.
Which jobs will AI displace? An analysis of 894 US occupations using a Delegation Resistance score built from O*NET work-context data and checked against 4.4 million Indeed job postings.
The follow-up, six months on. 2025 BLS employment data shows fifteen million routine, non-embedded jobs shrank while other knowledge work grew, tested against both Delegation Resistance and Tyler Cowen's two-factor model of AI-aided growth.
A language for card games that compiles into hidden-information environments for OpenSpiel, with automated checks that each player sees only what the rules allow. Built for multi-agent safety evaluations.
What I believed to be true about building software at a startup, as of June 2024.