I study how AI systems make reliable decisions when the information available to them is incomplete, corrupted, uncertain, or misleading.
My research focuses on reinforcement learning, AI evaluation, robustness, uncertainty, interpretability, and reliable decision-making under imperfect information.
Research
Observation Reliability in Reinforcement Learning
I study how reinforcement learning agents can assess whether individual observations are reliable and how reliability information should influence sequential decision-making under observation corruption.
AI Evaluation and Robustness
I am interested in how evaluation design, distribution shift, adversarial transformations, and limited observations affect conclusions about the reliability and behavior of AI systems.
Interpretability and Model Behavior
I investigate methods for understanding model behavior, including internal representations, behavioral mechanisms, and causal interventions.
Publications & Manuscripts
Manuscript in preparation
Observation Reliability in Reinforcement Learning
Reliability estimation and sequential decision-making when an agent's observations may be corrupted.
Selected Work
LLM Refusal Interpretability
Mechanistic interpretability study of refusal behavior in Qwen2.5-1.5B using activation analysis, linear probes, transfer tests, and causal interventions.
Open LLM Safety Fine-Tuning
QLoRA fine-tuning and leakage-safe evaluation of an open-source language model on WildJailbreak for safety-aware instruction classification.
Autonomous Voice Agent Evaluator
Autonomous Python voice agent for evaluating AI receptionists through realistic phone calls, evidence-grounded analysis, and assurance instrumentation.
About
I work at the intersection of machine learning research, AI engineering, and scientific computing.
My broader research interest is reliable decision-making under imperfect information: understanding what AI systems can infer from the evidence available to them, when that evidence should be trusted, and how uncertainty should affect decisions.
I am also the founder of TraceVox, where I work on infrastructure for AI assurance, evaluation, observability, and reproducibility.