Multimodal AI-based computer-aided diagnosis for gastrointestinal endoscopic biopsies (Seegene Medical Foundation). Multiclass and multi-resolution MIL, priority-aware mistake severity training, and decision making under uncertainty. Currently PM for a follow-up project (Seegene Medical Foundation): an LLM-based automated interpretation system for comprehensive blood tests, along with research on a separate NIPT system that predicts fetal risk from unreliable maternal blood signals.

Next-generation medical diagnosis system based on AI (Seegene Medical Foundation). Whole-slide segmentation, weakly supervised microscopy pathology classification, microscopy diagnosis team manager.

Safety monitoring model based on IoT for subminiature sensors (IITP). Time-series modeling and anomaly detection in sequential data.

My third consecutive project as PM with Seegene Medical Foundation. We analyze comprehensive blood test findings with an LLM and generate the interpretation report automatically. My main task is the AX of a complex rule-based diagnostic system into an LLM-based one, with the requirement that the generated findings stay faithful to the underlying lab values instead of drifting with the numbers.

A separate track that estimates fetal risk from measurements taken in maternal blood. The core problem is how to fuse signals when they cannot be trusted, and at which confidence level a result should be surfaced to the user, so the system needs reliability estimation, rejection of inconclusive cases, and outlier detection working together.

Based on digital whole-slide images (WSIs) of colon tissue specimens, we developed a diagnostic model tailored for non-malignant cases. In this project, our primary focus was placed on: (a) pre-processing massive WSI datasets, (b) building a high-fidelity model capable of expert-level collaboration with pathologists, (c) optimizing the model's diagnostic prioritization for specimens, and (d) enhancing the model's rejection capabilities for uncertain estimations. Serving as the project manager, I not only led the entire lifecycle but also spearheaded the technical solutions for areas (a) through (d), which ultimately resulted in the successful conclusion of the project and peer-reviewed publicatios.

This project aimed to classify gastric and colorectal biopsy samples exhibiting benign conditions versus dysplasia within digital pathology workflows. We pioneered approaches to efficiently train models on gigapixel Whole Slide Images (WSIs) while ensuring robust adaptation to newly integrated datasets. Notably, I served as the Project Manager (PM) and lead developer for a follow-up study focused on diagnostic models for noisy, microscope-based images. In my role as PM, I successfully translated Seegene’s business requirements into concrete technical specifications, steering the team to deliver high-impact engineering outcomes. Leveraging the insight that manual microscopic observations by pathologists function as time-series data, I formulated a novel distance metric based on Dynamic Time Warping (DTW). This work was successfully published at ISBI 2025. Our developed solution is currently deployed in clinical settings to cross-verify pathologists' diagnostic decisions.

Time-series modeling and anomaly detection for sequential sensor data in industrial safety monitoring. Specifically, I developed a deep learning-based methodology that detects the onset of fire from time-series sensor streams collected by subminiature IoT devices and raises an early alert, so that hazardous events can be recognized and communicated before they escalate.