📄 A paper on evaluating visual prompt-injection guardrails is accepted in the BEAM 2 workshop at ECCV 2026 as an Oral presentation
Beyond the Verdict: Evidence-Aligned Evaluation of Visual Prompt-Injection Guardrails.
About Me
I am the Head of AI at Tynapse, where I lead research on AI trust and safety for autonomous AI agents. Previously, I was a Staff Engineer at Samsung AI Center, building multimodal RAG applications for internal knowledge management. I received Ph.D. at Computer Vision Lab, Seoul National University under the supervision of Prof. Kyoung Mu Lee. I like discovering new insights from ambiguous problems and translating complex technical ideas into practical solutions.
Interests
- AI Trust & Safety
- AI Agents
- LLM / RAG
Education
PhD Electrical Engineering and Computer Science
Seoul National University
BEng Electrical Engineering and Computer Science
Seoul National University
Experience
Head of AI
TynapseLeading research on AI trust and safety, building real-time monitoring and governance solutions for autonomous AI agents.Staff Engineer
Samsung Electronics, DS Division, AI Center (*formerly SAIT, AI Research Center*)Responsibilities include:
- AI Agent (Jun 2024 – Sep 2025): Built a multimodal Retrieval‑Augmented Generation (RAG) application for internal knowledge management, improving efficiency in accessing distributed information across teams.
- AI ISP (Jan 2024 ‑ May 2024): Designed and trained a natural language‑based Image Quality Assessment (IQA) model, enabling intuitive and context‑aware image evaluations.
- AI ISP (May 2022 ‑ Dec 2023): Developed an efficient dual‑pixel autofocus model optimized for low‑light environments, improving focusing speed and accuracy.
Visiting Researcher
Google ResearchResponsibilities include:
- Conducted research on vision‑language models for controllable image editing guided by natural language instructions.
- Proposed a novel approach using conditional classifier‑free guidance to enhance editing fidelity and semantic alignment.
Research Intern
Snap Inc.Responsibilities include:
- Developed deep learning models for efficient and high‑quality video frame interpolation.
- Designed and implemented the SNU‑FILM benchmark dataset for analytic evaluation.
Education
PhD Electrical Engineering and Computer Science
Seoul National UniversityIntegrated Master & Ph.D in EECS.
Thesis: Test‑Time Adaptation Methods for Video Frame Interpolation.
Supervised by Prof. Kyoung Mu LeeBEng Electrical Engineering and Computer Science
Seoul National UniversityGPA: 3.48/4.30
Recent Publications
(2026).
Beyond the Verdict: Evidence-Aligned Evaluation of Visual Prompt-Injection Guardrails.
To appear at the ECCV 2026 2nd Workshop on Benchmarking Evidence-Aligned Multimodal Reasoning (BEAM 2) (Oral).
(2026).
Risk Is Not Review Value: Wrong-Answer Exposure Under Bounded Review Budgets.
KDD 2026 Workshop on Secure and Trustworthy Large Language Models (SeT-LLM) (Oral).
(2026).
Cross-Agent Campaign Attribution: Linking Asynchronous Attacks Across LLM Agents.
ICML 2026 Second Workshop on Agents in the Wild: Safety, Security, and Beyond (AIWILD).
(2025).
Stable Autofocus with Focal Consistency Loss.
Proceedings of the IEEE/CVF winter conference on applications of computer vision.
(2023).
Exploring positional characteristics of dual-pixel data for camera autofocus.
Proceedings of the IEEE/CVF International Conference on Computer Vision.
Recent News
📄 A paper on review-budget-aware LLM evaluation is accepted in the SeT-LLM workshop at KDD 2026
Risk Is Not Review Value: Wrong-Answer Exposure Under Bounded Review Budgets.
📄 A paper on cross-agent attack attribution is accepted in the AIWILD workshop at ICML 2026
Cross-Agent Campaign Attribution: Linking Asynchronous Attacks Across LLM Agents.
📄 A paper on stabilizing autofocus is accepted in WACV 2025 as an Oral presentation
Stable Autofocus with Focal Consistency Loss.
📄 A paper on autofocus is accepted in ICCV 2023.
Exploring positional characteristics of dual-pixel data for camera autofocus.