Projects
Selected research themes that I am interested in. For a full list of papers, see publications.
Scalable Delphi
LLM panels running the Delphi protocol achieve strong calibration against benchmark ground truth and align closely with human expert panels, reducing elicitation time from months to minutes.
Certified Training
Data poisoning is real and empirical defenses keep getting bypassed. We build methods that prove a model's predictions could not have been changed by any bounded poisoning of its training data.