Selected work

Personal · Production · 2023–PresentRV Badminton App서울의 한 배드민턴 클럽을 실제로 굴리는 4개 서비스 · 3개 언어 플랫폼. 단톡방 수작업 정산을 실시간 기록·리더보드와 재현 가능한 레이팅 엔진으로 바꿨습니다.#spring-boot#flutter#fastapi#duckdb#system-design DFPlantyNet · 2025–PresentOn-Device Deepfake Image DetectionI proposed and built an MLOps pipeline (Ray, DVC, Airflow, W&B) for an on-device deepfake detector, cutting an experiment from three days to eight hours and tracing a regression to its stage in an hour instead of four days.#mlops#ray#dvc#airflow#litert AIPlantyNet · 2025–presentAI Agents for Magazine Article ProcessingAn LLM + VLM system that summarizes, translates and classifies magazine articles and turns scanned pages into web-view articles, served through FastAPI endpoints and Dagster jobs. Three case studies show how I measured and improved its cost, recovery and quality.#langgraph#vllm#vlm#dagster#mlops RAGPlantyNet · 2025Multi-Agent Magazine RAG ChatbotI turned my first plain RAG chatbot into a router-based agentic RAG, and built a five-layer RAGAS + golden-set evaluation that made VOC processing, from user report to shipped update, go from two weeks to two days.#langgraph#milvus#ragas#agentic-rag RECPlantyNet · 2026 · Self-initiatedMagazine Recommendation SystemI proposed recommending articles by why people read them, not by what other users clicked. An LLM facet graph over the company's magazine archive hit readers' hidden interests 2.3× more often than the current recommender, and the proposal became an official project.#llm#knowledge-graph#neo4j#recommendation DVPlantyNet · 2026 R&DSpeaker-Conditioned Deepfake Voice DetectionThe plan was a generic deepfake voice detector on the phone. I measured it first, showed it missed about 80% of the latest cloned voices, and changed the R&D plan to asking "is this really the person you know?"#audio#anti-spoofing#on-device#r&d-planning