Overview
Worked as a research intern at the Ultimate Interface Lab, UNIST, advised by Prof. Seongkook Heo. Carried out my graduation research on BrowserLens, an interactive workspace for diagnosing web-agent trajectories, submitted to the UIST 2026 Posters track as joint work with Mingyu Kim and Seongkook Heo.
Technologies
Key Achievements
- Designed a layered state-action graph that reorganizes a recorded run around browser-state identity across three layers: normalized URL, viewport, and state-anchored action
- Built a review workspace where selecting a node retrieves the screenshots, DOM and ARIA snapshots, tool logs, and model context recorded at that state, so the graph indexes the original evidence instead of replacing it
- Implemented four visual cues — repeated node, loop edge, no-op action, and hidden-cue badge — supporting a cue → tentative failure point → evidence review flow, with no-op detection requiring unchanged normalized DOM and ARIA plus at most a 1% screenshot difference
- Demonstrated the workspace on a recorded Y Combinator directory run that ended with
exceeded max stepsafter 98 tool calls, where the URL layer surfaced a 35-visit company page and a circuit between company pages - Wrote the poster paper BrowserLens: Interactive Diagnosis of Web-Agent Trajectories with a Layered State–Action Graph and scoped the follow-up study comparing a linear step viewer, a viewer with linked evidence, and the full workspace