- 🔍 지능형 검색 — 키워드가 아니라 사용자의 의도를 읽어 텍스트 · 이미지 · 커머스를 가로질러 찾아 주는 검색을 연구합니다.
- 🧠 역할이 분명한 LLM 에이전트 — 모델에 맡길 일과 코드가 보장할 일을 나누고, 모델은 꼭 필요한 자리에서만 부릅니다.
- 🧪 점수보다 원인 — 어긋난 결과를 하나씩 따라가 원인을 가리고, 평가셋이 정답을 놓치면 평가셋부터 다시 만듭니다.
- 🎓 부경대학교 대학원 산업및데이터공학전공 석사 · TEAMLAB
| Year | Paper | Venue | Highlights |
|---|---|---|---|
| 2025 | SoFT · Soft Filtering: Guiding Zero-shot Composed Image Retrieval with Prescriptive and Proscriptive Constraints 1st author · Paper · Code |
AAAI 2026 Workshop Frontier in IR |
Plug-and-play reranking, no retraining CIRR R@5 +12.94 · CIRCO mAP@50 +6.13 |
| 2026 | CABIN · Controlling Augmentations for Balancing Invariance and Diversification in One-Shot Image Classification 3rd author · DOI |
Neurocomputing Vol. 699 |
LLM descriptor module for diffusion augmentation ResNet50 +7.52pp over SOTA on 6 benchmarks |
| 2023 | Dewy · A User Study of Prompt Engineering Techniques in a ChatGPT-based Skin Counseling Chatbot Lead · DOI |
Journal of Korea Innovation Society |
End-to-end chatbot (LINE · Flask · PostgreSQL) 50-participant randomized study |
|
2026 · AX Hackathon (MUSINSA track) · Solo
Explains how a garment will fit your body from product measurements alone, then edits the model photo into a reference image for your body shape. Works even for new items with zero reviews. Codex plugin 6-skill pipeline image editing
|
2026 · Fllow, influencer collaboration OS · AI Engineer
Finds real collaboration offers among fan messages and spam in Instagram DMs, anchors each offer to the message that started it, and keeps the project record up to date as the conversation continues. Gemini 5-stage LLM pipeline state machine Supabase
|
|
2024 · PKNU × Upstage Hackathon · Team Lead
Fashion search that starts from a situation or mood instead of a product name: an LLM composes a full outfit, and each item is matched to real products. Solar LLM CLIP hybrid matching
|
2023 · Genesis Lab · Semantic Search & RAG
Retrieval pipeline that lets AI meeting agents ground their arguments in a database too large for the context window — describing every table and column, then narrowing candidates to the meeting topic. RAG semantic search top-down + bottom-up
|
| 기간 | 소속 | 내용 |
|---|---|---|
| 2024.03 – 2026.02 | 부경대학교 대학원 · 산업및데이터공학전공 석사 | TEAMLAB에서 멀티모달 검색 연구 · 학점 4.43 / 4.5 |
| 2023.07 – 2024.06 | Genesis Lab · AI 연구팀 | 인턴 → 정규직 전환. LLM 에이전트 프로젝트를 기획 · 개발하고, ZUICY AI톡의 환각을 줄이기 위해 RAG와 대화 메모리를 설계 |
| 2020.03 – 2024.02 | 부경대학교 · 시스템경영공학부 학사 | 학점 4.03 / 4.5 |
- 🥈 우수상 · 농넷 농산물 가격예측 AI 경진대회 — 한국농수산식품유통공사 (2021.12)
- 🌊 대한원격탐사학회장상 · JOISS 해양과학 빅데이터 활용 경진대회 — 대한원격탐사학회 (2021.11)
- 🏙️ 입선 · 국토도시 데이터 분석 과제 — 한국토지주택공사 (2021.10)
세 건 모두 팀 가온(5명)으로 참가했습니다.


