Selected work I've built — AI benchmarks, memory evaluations, research systems, and developer tools.
Local, notebook-first editor for Jupyter .ipynb files with scoped AI-agent editing. The agent proposes cell changes; the backend validates and applies them — it never mutates your notebook directly. Per-turn Blocking (edit only the cells you scope) or Trusted (whole-notebook add/delete/reorder/edit). Runs on localhost.
Built a benchmark for measuring social cognition in frontier AI models across implicit norm detection, emotion recognition, cooperative breakdown detection, and group decision prediction. The benchmark uses Bayesian 2PL Item Response Theory to combine heterogeneous text and image tasks into a single ability score, keeping results comparable even when models answer only a subset of items.
Worked on LoCoMo evaluation for Continua’s long-term memory system, including benchmark setup, control-trace analysis, and failure-mode review. Analyzed how the benchmark design, scoring harness, retrieval traces, and answer-generation behavior affected results, helping separate memory-system performance from artifacts of prompts, judging, and evaluation setup.
A Streamlit analysis tool for clustering messy conversation topic exports. Analysts upload CSVs, choose subtopics or big topics, filter by cohort, group chat, or phone-number rules, then tune DBSCAN over all-MiniLM-L6-v2 embeddings to inspect topic clusters, compare populations, and export cluster summaries.
A zero-dependency Python tool for auditing coding-agent sessions: it reads Claude Code and OpenAI Codex logs and shows, task by task, which skills, tools, MCP calls, and subagents fired, plus which available skills never activated. Built for debugging the gap between available context and what an agent actually used.
A GUI application for correcting drift in eye-tracking data from reading studies — manual, semi-automated, and fully automated methods, plus filtering, visualization, metrics, and format conversion (EyeLink, ASCII, CSV, JSON). Published in Behavior Research Methods, 2025.
Co-designed, conducted, and analyzed two user studies on how people provide personal context to AI prediction systems. The surveys examined whether users could surface useful “human features,” how that process affected willingness to use the model, and how user-selected features compared with algorithmically selected ones. Published at ACM IUI Workshops, 2025.