Ph.D.
Identifying where lab-validated ML breaks in the field.
I am currently working on edge-cloud VLMs, quantizing the vision encoder for on-device deployment on a Qualcomm NPU, with a manuscript in preparation on where evaluation breaks between development and deployment. In parallel, I am extending my PhD research line into synthetic data generation for cross-environment ML transfer and explainable AI; a Google TPU Research Cloud proposal is under review. To validate the orchestration pattern for that work, I built a cost-efficient multi-agent LLM pipeline as a PoC.
In 2025, I completed my Ph.D. at Ghent University-imec, and was previously a RA at Academia Sinica. Both my PhD and MSc research were conducted on real-world data, covering 5000+ hours of uncurated audio streams and extended video across multiple deployment sites.
My research focuses on building solutions at the mechanism-level while under real-world constraints. Recognition includes 13 peer-reviewed articles (7 journals, ICIP oral), 150+ citations, APICTA award, and Top 3% master’s graduate honor.
- Deep Learning
- Computer Vision
- Audio Privacy
PhD in Computer Science Engineering
Ghent University, Belgium
MSc in Computer and Communication Engineering
National Cheng Kung University, Taiwan
BSc in Electrical Engineering
National Cheng Kung University, Taiwan
- 2026 to now
ResearchEdge-Cloud VLM: vision encoder quantization for on-device deployment
Quantizing a VLM vision encoder for a Qualcomm QCS8550 NPU. Eleven configurations evaluated on GPU, four actually executable on target. The paper is about where evaluation breaks between development and deployment. Manuscript to be submitted.
- 2026.06
ProposalGoogle TPU Research Cloud proposal, under review
Reality-Bounded Synthetic Data via Decoupling and Recombination: extending the PhD line to cross-environment event and behaviour transfer, and person-level stress-testing for deepfake detection.
- 2026.01
BuildMulti-agent LLM orchestration PoC
Cost-efficient routing, context isolation, and model gating, validated on a CPU/API-only setup before committing to GPU-heavy compute.
- 2025.11
MilestonePh.D. conferred by Ghent University-imec
From Lab to Street: transferable and privacy-friendly deep learning for urban surveillance, built on 5000+ hours of uncurated real-world audio.
- 2025.06
PaperSource-free model transferability assessment
Published in MDPI Sensors: ranking model transferability for smart surveillance without access to source data.
- 2025.01
PaperEmbedding-based pair generation for contrastive learning
Published in Frontiers in Robotics and AI: audio-visual representation learning on real surveillance streams.
