AI engineer and independent AI-safety researcher
AI engineer with more than twelve years of experience, now conducting independent AI-safety research on agent evaluation, social behavior, and mechanistic interpretability. Built production agentic systems, evals, and monitors.
Public experiment · 2026
Tested Claude Code across 20 Hangman games. The agent kept its secret word in 194 of 194 checks when private working memory carried between turns, compared with 47 of 177 checks when that memory was deleted.
NeurIPS 2026 workshop submission · Under review
Studied how misleading tool-output formatters can cause agents to reach false conclusions and evaluators to accept them. The paper’s audit prompt found an OpenClaw attachment-fetching bug; an OpenClaw maintainer merged the fix before submission.
NeurIPS 2026 workshop submission · Under review
Studied how absorbing basins can form in reinforcement-learning environments. Under the tested learning dynamics, models that began with betrayal tendencies reinforced betrayal, while models that began cooperatively reinforced cooperation.
GeBNLP at AACL 2026 submission · Under review
Tested whether concept-ablation fine-tuning could prevent a model from using a gender shortcut without retaining the ablation at inference. Replacing the concept direction with random noise produced higher out-of-distribution accuracy than replacing it with zeroes.
AI Engineer · Pro bono
Senior Software Engineer → Head of DevOps → AI Architect · Contract
Founder · Language-learning startup
Senior Software Engineer and Consultant
Research and ML
PyTorch, ablation, multi-seed evaluation, distribution-shift evaluation
Evals
Evaluators, eval sets, Promptim, LangSmith, and Langfuse
College of William & Mary