I’m an independent AI safety and evaluation researcher based in North Carolina’s Research Triangle. My work focuses on a methodological question: how can we build evaluation environments we can trust when models, agents, and evaluators themselves can introduce measurement failures?
Current work focuses on evaluation reliability, agent oversight, and adversarial evaluation, including Russian- and English-language settings.
- multi-agent-safety-sim — An experimental harness for studying multi-agent oversight and trace–report consistency through simulations, deterministic checks, and separately documented model runs.
- RuSafetyBench — Russian-language adversarial evaluation cases and qualitative pilot studies exploring how linguistic and contextual framing affects model behavior.
- Agentic Execution Guardrail Eval — A manually curated dataset for testing guardrails against prompt injection and risky agent-generated commands or code.
Simulated behavior and deterministic checks are distinct from empirical model evidence; project notes document the scope and limits of each.