Researcher · Engineer · Builder
I am an undergraduate student in Computer Science & Engineering at Konkuk University and an undergraduate researcher at UTL Lab, Korea University. My work focuses on trustworthy multimodal systems—evaluating AI safety guardrails, synthetic-video detectors, and the failure modes that appear beyond benchmark settings.
Previously, at DMS Lab, I studied deep reinforcement learning and large language models. I also enjoy turning research ideas into usable products, from a diffusion-based handwriting coach to a serverless iOS calendar.
My research interests include:
Selected work
Research questions and products I have taken from an initial idea to evidence.
Designed cross-editor evaluations for synthetic-video detectors across unseen editing models, compression, VAE reconstruction, and diverse edit concepts. The study revealed weak unseen-editor generalization and lower detectability for motion edits.
Built an evaluation framework for multimodal guardrail models, measuring over-refusal and joint image–text robustness while investigating systematic text bias in safety reasoning.
Created a personalized handwriting system that combines a writer's style with a structurally balanced reference using a conditional diffusion model and noise-prediction blending.
Shipped a deliberately simple iOS calendar centered on fast interaction and privacy-friendly device sync. The app uses iCloud synchronization without a separately operated application server.
Where I learned
UTL Lab · Korea University
Researching multimodal safety guardrails and the reliability of synthetic-video detection systems.
DMS Lab · Konkuk University
Studied deep reinforcement learning and large language models through implementation-driven research sessions.
Team Lett
Explored technical entrepreneurship by designing and building mobile and web products with a small team.
Konkuk University · Expected February 2027
Major GPA 4.24 / 4.5
How I work
I am most interested in systems whose failure modes can be measured, explained, and improved.
Guardrails that reason jointly over images and language without relying on shortcuts or excessive refusal.
Robust evaluation of detectors under unseen generators, editing concepts, compression, and reconstruction.
Practical uses of generative models and reinforcement learning that connect model behavior to human needs.
Technical toolkit
Contact
I am always happy to discuss research, engineering, or collaboration.
