M.S. student in Artificial Intelligence at Yuan Ze University (expected Jul 2027), working on deep learning for image and audio data with PyTorch. I like finding out how much performance survives when you take the compute away.
Eligible for Taiwan R&D Substitute Service (RDSS), available Jul 2027 · 研發替代役資格
Focus — deep learning · computer vision · object detection · speech separation Also build — Java / Spring Boot backends, containerised deployments
dinov3_detr — Resource-efficient object detection on a frozen ViT.
A DETR-style head with 8.5M trainable parameters on a frozen DINOv3 ViT-B/16, trained end-to-end on COCO 2017. Reaches 43.1 mAP on val2017 — 65% of the official DINOv3 detector's accuracy using 1.4% of its parameters, on a single consumer GPU.
PyTorch DETR Vision Transformer COCO MLflow
Denoiser — Two-speaker speech separation.
Conv-TasNet with a dual-path encoder and an input-adaptive slice-fusion mechanism (under 3K extra parameters). Reaches 14.29 ± 2.79 dB SI-SNR over 1,440 held-out utterances.
PyTorch Conv-TasNet Gumbel-Softmax SI-SNR
SyncSpace — Notion-style collaborative workspace.
Nested page trees, a block editor with drag-and-drop reordering, and 18 REST endpoints. Runs from a single docker compose up.
Java 17 Spring Boot JPA / Hibernate MySQL Docker
Training multi-label classification models on chest X-rays (RANZCR CLiP), and deepening the MLOps side of my work — model monitoring and CI/CD for training pipelines.