AI Engineer @ PlantyNet — on-device deepfake detection, agentic LLM systems, and the MLOps that keeps them running in production. M.S. in Computer Science (AI), Korea University · Advised by Prof. Seungryong Kim.

I work across two halves of the same problem: the research that makes a model good, and the systems that make it useful. Lately that means training and shipping on-device vision AI models, orchestrating multi-agent RAG agents over a 1,400-magazine archive, and building MLOps pipelines underneath both.

As an engineer I keep returning to the same three questions: why a system behaves the way it does, how an idea becomes something useful, and what to improve next. I write through those on my blog.

M.S. in Computer Science (Artificial Intelligence) · Korea University 2022.03 – 2025.02 Advisor: Seungryong Kim

B.S. in Statistics · Korea University 2018.03 – 2022.02

AI Engineer · PlantyNet — Seongnam, South Korea 2025.02 – present

• On-device deepfake detection shipped to Android via TFLite — trained on a large-scale corpus of real and synthetic faces, with a Korean-face subset added to close a domain gap.
• Research platform rebuild on Ray, DVC, Lightning, Hydra, and W&B, replacing a script-driven workflow and cutting both experiment turnaround and GPU usage.
• Multi-agent magazine RAG chatbot — a LangGraph routing orchestrator delegating to article-QA and recommendation agents over hybrid Milvus retrieval (bge-m3 dense + Korean BM25), with layered RAGAS evaluation separating routing, retrieval, context, and generation quality.
• Agentic content pipeline for OCR correction, summarization, translation, and web-view generation. An ExpeL-style experience memory raised processing success rates, and knowledge distillation into a lightweight classifier cut LLM inference cost.
• Knowledge-graph recommender (in progress) — LLM facet tagging into a Neo4j content graph, with Personalized PageRank candidate generation and contextual-bandit exploration.

Visiting Researcher · Queen Mary University of London — London, United Kingdom 2024.08 – 2024.11

Object-centric 3D reconstruction from monocular video — recovering per-object geometry and semantics with 3D Gaussian Splatting, then evaluating robot navigation and manipulation in the reconstructed scene.

Graduate Researcher · Korea University Computer Vision Lab — Seoul, South Korea 2022.03 – 2025.02

• Real-time audio-driven 3D talking heads
• 3D pose-conditioned diffusion for person re-identification
• pose-estimation/sensor-fusion indicators for dementia screening.
• TA for Samsung Electronics - internal courses on diffusion models and 3D reconstruction.

ML & Research

Lightning · Hydra · Weights & Biases · ONNX · TFLite / LiteRT

LLM & Agents

LangGraph · LangChain · vLLM · RAGAS · Langfuse · Arize Phoenix

Backend & Data

DuckDB · Milvus · Neo4j

MLOps

Docker Compose · Airflow · MLflow · Dagster · Ray Data · DVC · Triton

Infra & Observability

Mobile

Kyusun Cho, Joungbin Lee, Heeji Yoon, Yeobin Hong, Jaehoon Ko, Sangjun Ahn, Seungryong Kim

Jaehoon Ko, Kyusun Cho, Joungbin Lee, Heeji Yoon, Sangmin Lee, Sangjun Ahn, Seungryong Kim

Jaehoon Ko, Kyusun Cho, Daewon Choi, Kwangrok Ryoo, Seungryong Kim

Mira Kim, Jaehoon Ko, Kyusun Cho, Junmyeong Choi, Daewon Choi, Seungryong Kim