About
I like systems that make their reasoning inspectable.
My work sits between AI agent reliability, evaluation, data engineering, and applied machine learning. I care about what a system does when the clean path breaks—and whether its claims can be traced back to evidence.
From statistics to agent systems
I’m a Systems Engineering M.S. student at the University of Pennsylvania with a Statistics and Economics background from the University of Connecticut. That combination shapes how I work: I move between statistical evidence, software behavior, and the operational rules that connect them.
In my current research collaboration with UConn researchers, I study reliable tool-using agents: runtime intervention, recoverable multi-step execution, evaluation design, and closed-loop control. The central question is practical—how do we make an agent safer without losing the ability to finish useful work, and how do we evaluate what happens after an intervention?
Earlier work
University of Connecticut · Research Assistant
I integrated satellite imagery, OpenStreetMap building footprints, and UT-GLOBUS data for urban analysis. The work included spatial statistics, geospatial visualization, and an HPC/Slurm workflow for reprocessing building-level data across 413 major U.S. cities.
GBCS — SkyIT Services · Backend Developer Intern
I collaborated with frontend and backend engineers to deliver and release the Voop application for a client, developed and tested REST APIs with Express.js and Django REST, and supported a Firebase-to-MySQL migration through schema refinement and data-consistency checks.
Beijing QIANSHIDU Trade Co., Ltd. · Data Analyst
I analyzed more than 1 GB of two-year business data and developed a GPT-assisted demand-analysis workflow using SHAP and K-means. The analysis supported a shift in product allocation toward neighboring-country markets.
What I value
Reproducibility. A result should come with enough structure to be checked or rerun.
Clear boundaries. Course work, collaborative research, open-source extensions, and production systems should never be presented as the same kind of evidence.
Useful failure analysis. Negative results, blocked actions, stale state, and recovery paths often teach more than a polished demo.
Outside the work
I enjoy basketball, reading, writing, and spending time with Molly. Her corner of the site is intentionally personal—and stays separate from the professional portfolio.