• Video Content Moderation
  • Benchmark Dataset: Develop an end-to-end process for Trust & Safety benchmark dataset development, covering ① video indexing, ② pseudo-labeling, and ③ labeling tool development and management.
  • Performance Optimization: Analyze and evaluate video data to optimize model performance, continuously refining rubrics and pseudo-labeling methods.
• Video Understanding
  • Video Indexing: Build vLLM-based pipelines for efficient large-scale content processing and utilization, integrating new models to expand features.
  • Scene Segmentation: Research multimodal scene segmentation using visual, audio, and speech information for context-aware video understanding.

• Conducted research spanning video anomaly detection and medical image segmentation, with 4 first-author publications.
  • Anomaly Detection: Developed AnyAnomaly, a zero-shot VAD framework using LVLMs (WACV 2026); developed VideoPatchCore extending PatchCore to the video domain (ACCV 2024); proposed MAMA, an SSL-based framework to amplify anomaly signals via frame destruction (IEEE Access)
  • Segmentation: Developed DS-UNETR, a dual-stream fusion U-Net Transformer for 3D medical image segmentation (IEEE BigComp 2024)
• Graduation thesis: VideoPatchCore: An Effective Method to Memorize Normality for Video Anomaly Detection · Research Portfolio
• Advisor: Prof. Sanghyun Park (sanghyun@yonsei.ac.kr)

Poster presentation

Accepted at NVIDIA GTC 2026, San Jose, CA

• Enhanced video retrieval performance through scene-based video indexing
• Conducted Video QA experiments comparing uniform segmentation and scene-based segmentation on long-video benchmark

Team project

Grand Prize in Capstone Design Contest (Top project)

• Led the Fridge team; designed and implemented an end-to-end AI system
• Developed a vision-based AI chatbot for automatic food expiration-date recognition and management
• Conducted industry interviews with Korea E-mart (Shinsegae Group) to validate real-world adoption scenarios