• Backed by Y Combinator• Spearheaded development of LLM solutions for leading enterprises, including AngelList, Gusto, Veo Technologies, and Bull Moose Tube• Built first LLM Agentic marketing platform

• Our paper "CoRAL: Collaborative Retrieval-Augmented Large Language Models Improve Long-tail Recommendation" has been accepted by KDD 2024

• Reconstructed real-time CNN for image 2x upscaling to improve debris clarity; increased ~4db PSNR above base model• Converted model to ONNX format and deployed it on Linux server; attained upscaling time to 30ms per frame• Refactored inference code and streamlined data transmission with Pytorch tensor operations, boosting 30% runtime speed• Built self-learning Docker pipeline for feature database matching and updating through YOLOv8, CLIP-ViT feature extractor, DreamSim perceptual metric, and DeepSORT tracking algorithm

• Worked alongside Prof. Sharon Di to research human mobility prediction in NYC during COVID-19 pandemic period based on graph neural networks• Outperformed SOTA methods via reducing RMSE by 28% and MAE by 46%

• Worked on AI-Care product development and presented it at The Cares Taipei 2021, international exposition and conference focused on healthcare industry• Fine-tuned PoseNet for fall detection utilizing data augmentation in PyTorch DataLoader, beating baseline by 5% • Established custom streaming pipeline leveraging NVIDIA DeepStream for array input to RTSP output• Flashed NVIDIA Jetson kernel into eMMC, boosting 18x reading speed and 5x writing speed over SD card

• Implemented contactless heart rate measurement for multi-person detection by leveraging rPPG algorithm and face recognition models; increased system’s FPS by 1.5 times via implementing face tracking algorithm• Engineered multi-angle face recognition through ArcFace and 2DFAN; achieved system performance up to 20 FPS

• Constructed deep learning-based OCR models in PyTorch to recognize defective printed characters on steel plates; improved overall identification accuracy by 6% and boosted transfer learning efficiency by 7%• Integrated real-time OCR system on NVIDIA Jetson AGX Xavier by Basler industrial camera and Socket API, reducing production time across various factory lines by 60%