Building a science of interpretive work across humans, autonomous agents, and their collaborations.
I am currently a Malone Postdoc Fellow at Johns Hopkins University, where I am mentored by Mark Dredze, Ziang Xiao, and Chien-Ming Huang. My research helps people make sense of large datasets and develop intuition from the insights they reveal. I use a range of social science methodologies to support this process. I also bring human perspectives into AI systems to make them more understandable, interpretable, and intuitive. My work is interdisciplinary, spanning AI, HCI, software engineering, and the social sciences. I publish my work in venues such as NeurIPS, EMNLP, CHI, TOCHI, and ICPC. Alongside my research, I am the lead organizer of the first Interpreting Agent Behavior (IAB) Workshop@NeurIPS 2026 and a co-organizer of the LLMs as Research Tools workshop@CHI 2024.
Organizing the IAB Workshop @NeurIPS 2026, on bringing interpretation methods from HCI and the social sciences to build interpretability for autonomous agents.
Gave a talk, “Toward Seeing What Agents Do: Interpretive Work in the Age of Agentic AI,” at Monash University
Gave a talk, “Toward Seeing What Agents Do: Interpretive Work in the Age of Agentic AI,” at Fudan University (News)
Our preprint How to Interpret Agent Behavior is out! We introduce Act*ONOMY, a hierarchical taxonomy for analyzing autonomous agent behavior at runtime.
Our CodeMap paper received an award at ICPC 2026 (a CORE A conference in code comprehension)! ACM SIGSOFT Distinguished Paper Award
Happy to give a talk at Advanced HCI at JHU (Slides)
Gave a Claude Code How-To Session (Slides) for 40+ JHU researchers (Master's students, PhD students, postdocs, and faculty): covering practical usage and best practices of Claude Code and OpenClaw, provided live demos.
I attended VL/HCC 2025 conference in North Carolina. Met new friends and had a great time!
I attended CHI 2025 conference in Japan, and met a lot of old and new friends!
We introduce mindcoder.ai, a web application designed to support automated qualitative analysis. The user only needs to follow three steps to obtain qualitative analysis results-1) upload data, 2) start coding, and 3) get analysis report! See how to use-video.
Started my postdoc fellowship at Johns Hopkins University!
We have launched CollabCoder Website. Construction and refinement is ongoing!
One first-authored paper "CollabCoder" and one coauthored paper "Help Me Reflect" have been accepted to CHI2024! (Preprints will have minor refinements in final versions.)
How can AI support, rather than replace, the interpretive work of reading unstructured text? I design and evaluate systems that help people code qualitative data, align interpretations across multiple coders, and develop theories from raw text, while keeping human judgment central.
Real-world codebases are messy, and newcomers often struggle to build an accurate mental model. I build tools that help developers understand unfamiliar code, review AI-generated changes critically, and collaborate with AI assistants without giving up their own judgment.
As AI systems take on increasingly autonomous roles, the bottleneck shifts from "can they do it" to "can humans understand or control what they do." I study how humans can understand and control agents when they run autonomously.
A human-AI collaborative tool for code comprehension. Helps developers understand complex codebases through interactive visualization and AI-guided exploration.
An AI-powered platform for flexible qualitative data analysis. Supports open coding, sub-theme grouping, and theme generation with human-AI collaboration.
A GPT-powered workflow for collaborative qualitative analysis. Enables multiple researchers to code data together with AI assistance, improving inter-rater reliability and efficiency.
