Can an AI model be trusted once it leaves the lab?
Uncertainty and explainability are the trust axis; edge, low-resource, and challenging environments are the deployment axis. The intersection stays mostly empty, because uncertainty is studied on well-resourced models and efficiency is measured by accuracy alone.
That intersection is my research: trustworthy AI under deployment constraints. I build systems that stay calibrated and explainable on the hardware and in the conditions where they actually run, with applications in medical imaging and aerial and drone imagery.
Trustworthy AI under deployment constraints, applied to medical imaging and aerial and drone imagery.
Whether a model knows when it is wrong, and whether it can say why. Calibration and conformal prediction, separating epistemic from aleatoric uncertainty, out-of-distribution detection, and selective prediction that defers to a human rather than guessing. On the explainability side, faithful explanation: concept-based and language-grounded reasoning that can be checked against the evidence, rather than post-hoc saliency that merely looks convincing.
Whether a model works at all where it is needed. Efficient architectures for surgical carts, endoscopy systems, onboard drone processors, and consumer devices, where on-device inference also keeps sensitive data from leaving the site. Alongside this, robustness to the conditions real data arrives in: scanner and stain variation, motion blur, occlusion, smoke, altitude and viewpoint change, and scenes never represented in training.
A consumer-electronics-oriented agentic AI framework for polyp segmentation and clinical decision support, designed to operate across hospital workstations, mobile devices, and patient-facing consumer electronics.
EMCAF-Net, an Edge-Aware Multi-Scale Contextual Attention Fusion Network for accurate cell nuclei segmentation in cancer diagnosis, integrating edge-aware multi-scale fusion for boundary localization with a multimodal LLM for automated diagnostic reasoning.
A physics-informed neural network framework for the coupled chemo-mechanical PDEs governing lithium transport and stress evolution in battery electrode particles, replacing traditional finite element methods. Anisotropic, SOC-dependent diffusivity tensors derived from molecular dynamics simulations are embedded as inputs, enabling accurate modelling of irregular particle geometries from X-ray tomography data.
Full list on Google Scholar ↗
Division of Electronics and Electrical Engineering
Dongguk University
Seoul
South Korea
Google Scholar · GitHub · Kaggle · LinkedIn
I am looking for motivated graduate students interested in computational pathology, surgical AI, agentic AI systems, and physics-informed machine learning. If you are interested in joining my group, please send your CV and a brief research statement to my email.