About
I am an Applied Scientist II on the Agentic AI team at AWS, working on agent evaluation, agentic memory, long-horizon reasoning, and multimodal language models for Amazon Bedrock AgentCore and Strands Agents. I build evaluation LLMs and meta-evaluation frameworks that automatically assess and validate agentic systems. Previously, I worked on multimodal embedding retrieval for Amazon Lens Live on the Core Search team. I completed my Ph.D. in Computer Science and Engineering at POSTECH, advised by Prof. Suha Kwak. During my Ph.D., I collaborated with Ivan Laptev. I also interned under and closely worked with Douglas Gray at Amazon.
My research sits at the intersection of information retrieval and agentic AI, with one aim: building agents we can trust. Retrieval keeps what agents know accurate and grounded in evidence, and I care most about reliability, developing evaluation that measures and validates whether agents truly behave as intended. I first built search for people; now I build agents that retrieve, reason, and act over knowledge on their own, along with the evaluation that verifies they can be trusted at scale.
Experience
-
Applied Scientist II — Agentic AIAmazon Web Services · Bedrock AgentCore & Strands Agents- Building evaluation LLMs and meta-evaluation frameworks for LLM-as-a-Judge assessment of agentic systems.
- Research on agent orchestration, memory, context management, and long-horizon reasoning.
- Evaluated coding agents for Kiro, AWS's autonomous coding agent, including SWE-bench-style benchmarking.
-
Applied Scientist II — Core SearchAmazon.com · Visual Search (Amazon Lens Live)- Multimodal embedding retrieval and result verification for Amazon Lens Live.
- Reasoning-driven re-ranking in an MLLM-based contextual retrieval pipeline.
-
Applied Scientist Intern — Core SearchAmazon.com · Manager: Douglas Gray
- Generative retrieval for universal multimodal search, published as GENIUS (CVPR 2025).
Publications
International Conference
-
-
-
-
-
FREST: Improving Robustness of Semantic Segmentation via Source-free Domain Adaptation with Feature RestorationECCV 2024 -
PromptStyler: Prompt-driven Style Generation for Source-free Domain GeneralizationICCV 2023 -
HIER: Metric Learning Beyond Class Labels via Hierarchical RegularizationCVPR 2023Qualcomm Fellowship Finalist -
-
-
-
-
Embedding Transfer with Label Relaxation for Improved Metric LearningCVPR 2021BK21 / IPIU Best Paper -
-
Deep Metric Learning Beyond Binary SupervisionCVPR 2019Oral · 5.58% · 140+ citations
Preprint
Technical Blog
Education
-
Ph.D., Computer Science and EngineeringPohang University of Science and Technology (POSTECH)- Advised by Prof. Suha Kwak. Dissertation: Towards Retrieval at Scale via Compact Embeddings and Generative Modeling.
- Graduated with the highest distinction across the CSE, GSAI, EE, and CITE departments (press).
-
B.S., Undergraduate StudiesDaegu Gyeongbuk Institute of Science and Technology (DGIST)
Honors & Awards
- Graduated with the highest distinction across the CSE, GSAI, EE, and CITE departments (2025) · press
- Qualcomm Innovation Fellowship Korea (2024) · Winner — Efficient and Versatile Robust Fine-Tuning of Zero-shot Models
- Google PhD Fellowship Program (2023)
- BK21 Best Paper Award (2023) · Winner — Self-Taught Metric Learning without Labels
- BK21 Best Paper Award (2023) · Winner — Combating Label Distribution Shift for Active Domain Adaptation
- Qualcomm Innovation Fellowship Korea (2022) · Winner — Self-Taught Metric Learning without Labels
- Qualcomm Innovation Fellowship Korea (2022) · Winner — Combating Label Distribution Shift for Active Domain Adaptation
- BK21 Best Paper Award (2022) · Winner — Embedding Transfer with Label Relaxation for Improved Metric Learning
- IPIU Best Paper Award (2022) · Gold Prize — Offline Active Domain Adaptation
- CVPR Outstanding Reviewer (2022)
- ICT Paper Contest (2021) · 2nd Place — Deep Metric Learning Beyond Binary Supervision
- SKT AI Fellowship (2021)
- POSTECHIAN Fellowship (2021)
- IPIU Best Paper Award (2021) · Grand Prize — Embedding Transfer with Label Relaxation for Improved Metric Learning
- Naver Ph.D. Fellowship (2020)
- Qualcomm Innovation Fellowship Korea (2020) · Winner — Deep Metric Learning Beyond Binary Supervision
Professional Service
- Conference ReviewerCVPR, ICCV, ECCV, ICLR, ICML, NeurIPS, AAAI, BMVC, ACCV, WACV
- Journal ReviewerIEEE TPAMI, IJCV, IEEE TIP
Patents
- Rehabilitation program creation method for muscle treatment and providing apparatusKR101648638B1, Republic of Korea