projects
Personalized agents, models of human behavior, and the systems behind them.
Can Personalized Agents Predict Social Experience?
Research collaboration with Kazuki Kawamura
- Developed personalized LLM agents that use adaptive interviews to model social preferences and assess connections based on users’ goals and context.
- Built a Python/PyTorch recommendation engine that learns from sparse pairwise scores to rank candidates and construct degree-constrained social graphs.
- Designed a longitudinal field study comparing feedback-updated and interview-only agents on compatibility predictions for unseen people. Planned measures: agreement with human judgments, match quality, LLM query efficiency, and interest in reconnecting.
User and Relationship Modeling from Conversation Data
- Built a system to model individual behavior and relationship dynamics from WhatsApp conversation data.
- Generated behavioral profiles linked to supporting messages; incorporated user corrections and tracked changes in expressed emotions and communication patterns.
- Compared users’ stated preferences with observed conversational behavior and interaction feedback.
Fine-Tuning Language Models to Predict Human Responses to Interventions
- Fine-tuned Qwen3.8-27B on SocSci210 with LoRA; compared standard SFT with additional objectives for matching response distributions within experimental conditions and predicting effects between conditions.
- Evaluated generalization to held-out studies using response-distribution and experimental-effect prediction errors; conducted ablations to isolate each additional loss term’s contribution.
The Registry
Semantic Search Engine
- Built an end-to-end Python pipeline for natural-language search over 18,611 university student profiles: data processing, embeddings, FAISS/HNSW approximate nearest-neighbor search, cosine similarity, and top-k retrieval.