About

From the Wrestling Mat to Building Useful Technology

As a former ssireum (Korean wrestling) athlete, I made a late but determined leap into computer science at Pusan National University, driven by a desire to create technology that serves real industries and everyday life — technology I could call my own.

Yet the undergraduate curriculum, centered on theory from dated textbooks, couldn't quench that thirst. I threw myself into internships and startups, where hands-on projects exposed the limits of my knowledge and made the need for deeper learning painfully clear.

Parsing Taught Me to Think Structurally

Searching for answers, I joined the lab of Professor Hyuk-Chul Kwon — one of Korea's first AI Ph.D.s and an authority on parsing — first as an undergraduate researcher and then as a master's student. Over two and a half intensive years of dependency parsing and neural-symbolic modeling, I learned to view language through relationships and dependencies rather than as flat sequences.^[1][2]

Combining learned models with explicit linguistic knowledge gave me a habit of thinking structurally that still shapes how I approach every problem.

Human-Inspired AI Taught Me What Structure Should Serve

Parsing taught me to think structurally; human-inspired AI taught me what that structure should ultimately serve.

During my master’s research, parsing trained me to view language through relationships and dependencies rather than as flat sequences. After joining Korea University’s Human-Inspired AI Research, working with Prof. Heuiseok Lim and Prof. Jaehyung Seo, my perspective broadened. Through industrial QA, foundation-model adaptation, and multimodal systems, I repeatedly saw technically strong models lose practical value when they ignored conditions, flattened meaningful organization, lost provenance, or failed to preserve information across interactions.

Since then, I have approached research by starting from problems that matter in actual use, identifying the technical bottlenecks that limit their utility, and asking whether solving those bottlenecks reveals a broader scientific problem. Explicit structure remains one of my primary tools—but the goal is not structure itself. The goal is to make increasingly capable AI systems genuinely useful to the people who rely on them.

From Real-World Needs to General Research Questions

My work follows this pattern. StyleDFS began with an industrial need: domain question answering over power-plant documents under on-premise and privacy constraints, which led to a domain-specific RAG framework.^[3] In an insurance QA pilot with Samsung Fire & Marine, conditions, exceptions, and provenance were easily lost during retrieval.^[4]

Those bottlenecks became general research questions. MultiDocFusion and M3DocDep preserve the hierarchical and cross-page organization of long documents;^[5][6] HiKEY and PILAR select and ground the evidence a system answers from;^[7][8] ADMIT controls what agents retain in memory across interactions;^[9] and my current long-video work extends this from spatial document structure to temporal information.^[10]

To pursue these questions beyond what a single researcher could address, I founded KUDoc, the Document AI research group at Korea University, on my own initiative. Together we have published 15 papers (including 6 first-author papers at NeurIPS, EMNLP, ACL, and CVPR) and completed 4 major industry–academia projects.

Human-Inspired AI for Multimodal and Agentic Systems

As AI becomes more multimodal, autonomous, and persistent, how do we ensure its capabilities continue to create value for people?

This is the question I want to pursue in my Ph.D.: how multimodal and agentic systems should represent, select, update, and verify information so that increasingly capable AI remains useful and reliable over longer interactions.

With the tenacity I learned on the wrestling mat, I intend to keep turning real human needs into research questions—and research into systems people can rely on.