AI assists.
People decide.
I build the parts that make AI trustworthy once it's live.
ctrlrun
A person approves the exact action, each effect runs at most once, and every action leaves a receipt.
Evals
Golden test sets, LLM-as-judge and human review for factuality and hallucination, before every release.
OWASP
Section author on agent audit trails for the GenAI Security Project.