Incoming AI major at UC San Diego's Jacobs School of Engineering. I build detection and reinforcement-learning systems for chest X-ray, endoscopy, and video understanding.
I work at the intersection of computer vision, reinforcement learning, and clinical imaging — mostly asking how models can localize disease, not just classify it. My first-author work built a text-guided region-region similarity architecture for chest X-ray abnormality detection, done in collaboration with Oxford-affiliated researchers and published in AJSR.
At MIT Lincoln Laboratory's Beaverworks Summer Institute (Medlytics), I worked on zero-shot brain tumor detection. I've since presented at UCSF AI Research Day and continue building research ties there.
Outside the lab: a black belt in Kung Fu under the International Wushu Sanshou Dao Association, AP Scholar with Distinction, and time spent mentoring and tutoring other students in AI.
Building a multi-box detection pipeline for endoscopic lesion analysis by combining a YOLO-based detector with MedGemma-derived supervision. The work is currently in active training and evaluation, focused on developing a clinically relevant system for medical video understanding rather than a one-off demo.
Built a demo-ready pneumonia classification workflow for chest X-rays using MedGemma, packaged as a Streamlit application with reusable preprocessing and inference logic for DICOM inputs. The project is designed to demonstrate how multimodal foundation models can be applied to medical imaging in an accessible, interview-ready format.
Region-region similarity based text-guided object detection for abnormality localization in chest radiographs, combining textual descriptions with visual features to improve diagnostic accuracy.
Led and contributed to a research-focused medical imaging effort through MIT Lincoln Laboratory’s Beaverworks Summer Institute, prototyping and developing a zero-shot framework for brain tumor analysis and exploring how foundation models can support diagnostic workflows across multiple tumor classes.
Through the Beaverworks Summer Institute and Medlytics track, I built and presented a set of rapid-turn medical AI projects that connected machine learning with real clinical problems. These included mammogram-based breast cancer classification using CNN and transfer-learning approaches, hypothyroidism prediction from clinical data with decision trees and MLPs, and sleep-stage classification from multimodal biosignals using 1D convolutional networks. Each project sharpened my ability to move from an idea to a functioning diagnostic pipeline under time pressure.
A CV pipeline detecting mask compliance from images — an early exercise in applying detection models to public health monitoring.
Image classification for identifying disease in tomato plants from leaf photos — applying the same detection instincts outside medicine.
Building UlcerDetRL's multi-box detection pipeline and exploring V-JEPA's latent prediction error as a free supervisory signal for label-efficient temporal action segmentation — alongside continued conversations with UCSF researchers on future collaboration.
Open to research collaborations in medical imaging, RL, and detection.