Value-Aligned AI Scheduling Assistant

Published:

Designing an evolving Bayesian model of a person’s priorities that integrates conversational, behavioral, and reflective evidence while preserving uncertainty. Built a grounded LLM pipeline (interviewer, evidence extractor, reviewer) so inferred values stay traceable to exact user statements. Introduced counterfactual value previews, letting a person see how a candidate action would shift the model’s interpretation before committing to it.

Ongoing work with Prof. Edith Law, Augmented Intelligence Lab, University of Waterloo.