Career Profile
Hello! I’m Thomas, a graduate student at the University of Michigan, Ann Arbor. Currently, I am pursuing a Master’s degree in Electrical and Computer Engineering under the Computer Vision track. My research interests lie in Computer Vision applications, particularly Autonomous Driving, as well as Contactless Patient Monitoring in the healthcare sector. Beyond academia, I have a keen interest in the financial markets and actively participate in investing and stock trading. As a car enthusiast, I enjoy delving into automotive mechanics, spending my free time working on cars, and enhancing my practical engineering skills.
Experiences
- Improving upon OpenEMMA, an End-to-End Multimodal model for Autonomous Driving by incorporating a retrieval-augmented prompting framework to fuse prior road knowledge with on-board camera feeds.
- Optimizing and accelerating computation speeds of the LLM to achieve offline capability operating on vehicles locally.
- Implemented and evaluated a Temporal Shift Convolutional Attention Network for physiological signal modeling, achieving high generalizability across datasets (Pearson correlation: 0.989, R² = 97.8%).
- Currently designing an Anomaly Detection pipeline for driver drowsiness by improving upon the LLaVA Computer Vision Transformer model.
- Developed and trained a custom 1D convolutional neural network (CNN) to identify informative bitplanes from RGB video data, improving heart rate prediction accuracy by filtering out noise in facial blood flow signals.
- Engineered a image processing pipeline to extract and structure 24-bitplanes (8 per RGB channel) from 93 training videos, enabling efficient ML model input for physiological signal enhancement.
- Integrated CI/CD pipelines into a Warehouse Management System written in Java to streamline deployments and improve supply chain efficiency by 20%.
- Ensured the efficiency and correctness implementation of the pipeline by using JUnit for unit testing before deployment.
Projects
Kaggle ISIC 2024 Competition - Skin Cancer Detection with 3D-TBP
- • Designed and trained a convolutional variational autoencoder to synthetically generate malignant skin lesion images, addressing class imbalance in a skin cancer classification task using latent space sampling.
Computer Use Automation
- Developing an LLM-powered desktop automation agent that integrates voice input, visual perception (OCR/screen parsing), and mouse/keyboard control to perform real-world GUI tasks autonomously.