3+ years building AI systems across medical imaging, 3D vision, time series, and generative data — for teams in Korea, Singapore, Switzerland, and Taiwan.
I take AI projects from an unsolved real-world problem to a working, measured system — usually in resource-constrained environments, often as the only engineer on the project.
Supported planning and operations for the mammography analysis AI program — data pipelines via AIHub, annotation tooling, plus a separate fracture-classification model's training/validation data QC.
→ Coordinated FDA approval process and Korea–Taiwan technical communication.
Built a workflow combining object detection, video inpainting, 3D reconstruction, and depth-aware insertion — replacing real objects in footage (e.g. a laptop) with 3D models.
→ Custom pipeline outperformed commercial tools (Luma Labs, Metashape) on the team's specific use case.
Built and fully Dockerized a time series prediction model trained on satellite data, deployed as a self-contained, reproducible environment.
→ Delivered as a remote AI consultant, end-to-end from model to deployment.
Built an AI model predicting sperm chromosome abnormality from motion-pattern data, supporting diagnosis of unexplained male infertility.
→ Helped secure ₩300M in government AI Voucher funding. Published at ICGHIT 2023.
Built an AI scoring model for the Beery-VMI visual-motor integration test. Used VQGAN to generate synthetic test images, solving a severe data scarcity problem for this clinical population.
→ Published at Korean Society of Rehabilitation Robotics 2022 and ICGHIT 2023.
Applied a GAN to amplify a CAD-drawing dataset for an elevator-control PLC auto-generation system, then trained a PyTorch classifier with incremental learning on the expanded set.
→ Increased the dataset 10× — from 700K to 7M records.
Used QGIS and GeoPandas to engineer spatial features from crime data, then trained a LightGBM model to predict high-risk areas for the Smart Policing Seoul Center initiative.
→ Achieved 80% prediction accuracy.
Statistically analyzed self-measured skin data from 200 users to build a scoring method, then delivered personalized product recommendations via a Firebase push-message service.
→ Published at ICGHIT 2022.
Problem: Classic supply-chain ordering (the "beer game" / bullwhip-effect problem) gets harder when multiple items compete for one shared budget — standard Q-Learning has no way to respect that constraint.
Approach: Modeled ordering as an MDP (state = inventory/backlog, action = discrete order quantities, reward = −total cost) and trained with Q-Learning. Added OptLayer — a constrained quadratic-programming layer originally built for robot-arm reaching tasks (Pham et al., 2018) — repurposed here to project any budget-violating order onto the nearest feasible one.
Result: Benchmarked against plain Q-Learning, a classic Lagrangian heuristic, and the textbook EOQ formula. Q-Learning + OptLayer achieved the lowest total cost of all four approaches.
Studio2J is a proxy-buying service I run for Korean illustrator stationery, tracking live markets across Seoul and Tokyo fairs. I'm building a RAG chatbot trained on real illustrator data I've collected — customers describe what they like, and it recommends matching illustrators from an actual, curated dataset rather than a generic catalog.
I'm always happy to talk about AI engineering, research, or potential collaborations. Reach out through any of the channels below.