About Me

AI engineer and photographer based in Taipei, Taiwan.

I'm a Computer Science and Engineering graduate of National Taiwan Ocean University (Class of 2026). RiffNode — a guitar effects studio built entirely with native Apple frameworks — won the Apple WWDC26 Swift Student Challenge.

At ITRI I trained and deployed multi-task driving-perception models down to a Jetson Nano. I led my capstone team's volleyball analysis system, which won the department's Software Development Silver Award, and I have kept building it as a research project — from one camera to the court in metres, tracked players and 3D ball flights. Lately I have been auditing whether guitar-tone models quietly change a player's touch, and building a macOS app whose architecture the compiler enforces. I like working across the stack — on-device ML, audio DSP, systems Swift — and turning hard problems into things people can actually use.

Experience

Teaching Assistant — Introduction to iOS Application Development

National Taiwan Ocean University

Department of Computer Science and Engineering

Keelung, TaiwanSep. 2026 – Jan. 2027

AI Image Recognition Intern

ITRI (Industrial Technology Research Institute)

Electronics and Optoelectronics Research Laboratories

Hsinchu, TaiwanJun. 2025 – Aug. 2025

·Trained YOLOPX, a multi-task panoptic driving perception network doing object detection, drivable-area segmentation and lane segmentation in a single forward pass — reaching 0.972 mIoU (drivable area), 0.793 mIoU (lane) and 0.838 mAP@0.5 over 4 classes after 200 epochs on ITRI's in-house driving dataset
·Extended the open-source YOLOPX codebase by ~1,900 lines (6 files rewritten, 5 added) for ITRI's data format, building a YOLO-segmentation loader and a Mosaic/MixUp/Albumentations augmentation pipeline to replace one that handled BDD100K only
·Owned the edge deployment path: network pruning, ONNX export, a tool resolving dynamic convolution-bias subgraphs that blocked TensorRT conversion, a TensorRT/PyCUDA engine for NVIDIA Jetson Nano, and image + detection streaming to a desktop PC over Ethernet/ROS at 30 FPS with sub-50ms latency
·Delivered a command-line batch-inference tool for ARTC (Automotive Research & Testing Center) running the trained model across image and video datasets including a new rainy-weather set; separately reproduced and benchmarked YOLOP, YOLOPv2, YOLOPX and A-YOLOM to identify deployment-viable designs