Research

An evaluation framework built on 13,000+ real user messages from psychological distress contexts. DelusionEval surfaces context-dependent failure modes in LLM safety behavior as conversation history evolves. These failures are invisible to single-turn benchmarks.

Evaluating digital behavioral interventions on adolescent screen time using longitudinal smartphone logs and mixed-effects models, with trajectory clustering to identify which psychological profiles predict meaningful vs. null change.

Constructing high-resolution behavioral time series from 100M+ smartphone screenshots and modeling depressive symptom trajectories with longitudinal CES-D assessments. Findings highlight behavioral precursors of mental health deterioration, with substantial individual heterogeneity.

Analyzed 10,000+ online citizen complaints (SeeClickFix, Google Maps) using NLP for sentiment and topic insights on infrastructure concerns like parks, street lights, and roads. Final deliverable combined literature review and empirical findings in urban data analytics.

NLP and statistical analysis of 800+ U.S. commencement speeches (1890–2020), tracing temporal shifts in rhetorical patterns, sentiment, and thematic structure across more than a century of public speech.

Co-authored a paper on how political conflicts and compromises can be expressed using simplicial algebraic topology, building on homology theory and current topics in algebraic topology research.