Research portfolio2025—Now
Selected research spanning generation-time safety signals, multi-turn jailbreaks, and reliable retrieval for high-stakes domains.
View case studies01
Selected research
LLM Safety · Evaluation
A process-level evaluation framework that traces refusal, compliance, and harmful-token signals across generation instead of reducing safety to a final-answer score.
Applied AI · Retrieval
A graph-grounded retrieval and reasoning framework for interpreting financial security regulations with more reliable domain evidence.
Adversarial ML · Jailbreaks
An analysis of how attacker-controlled context accumulates across turns and changes model behavior through incremental memory injection.
Complete dossier