Zhien Lai

M.S. Computer Science student at USC with experience in visual recognition, deep learning, reinforcement learning, and LLM-based agent systems. I am now particularly interested in game programming, interactive systems, and the integration of AI into games, with the goal of building technology that people can directly see, interact with, and experience.

• Graduation Project, Reinforcement Learning-Based LLM Fine-Tuning and Automatic Annotation System, Xi’an Jiaotong University | Spring 2026, Developed an end-to-end LLM training and data augmentation system integrating supervised fine-tuning, reinforcement learning, and automatic pseudo-labeling, Fine-tuned Qwen3-4B, Llama3.2-3B, and DeepSeek-R1-Qwen-1.5B and implemented PPO-based reinforcement learning with a multi-dimensional reward function, Built an automated Self-Instruct and LLM-as-a-Judge pipeline for generating and filtering pseudo-labeled training data, together with an interactive system for model management, training configuration, and real-time monitoring
• Special Study: Computer Vision, University of Notre Dame | Spring 2025, Developed an end-to-end image classification system, Utilized OpenCV for data cleaning, including Normalization, Gaussian blurring, and GrabCut for foreground segmentation and combined color and grayscale feature data augmentation to enhance input quality, Built a 9-layer convolutional neural network by TensorFlow/Keras framework
• Operating System Practice on Developing Ext2 File System, Xi'an Jiaotong University | Spring 2024, Developed a complete Ext2 file system based on C language and Huawei Cloud platform, enabling a full cycle of “create-mount-read/write-delete” functionalities, Established the whole disk layout, defining the superblock, group descriptor, inode, and data block bitmap
• Game Development with C Programming, Xi'an Jiaotong University | Summer 2023, Developed a mini-game based on C language and the EasyX graphics library, responsible for the entire development process, Achieved dynamic ball rebound through collision detection and mirror reflection algorithms, used non-blocking input polling and timed refresh mechanisms to ensure synchronization between game logic and animations, Utilized a modular design for the game architecture, incorporating random initialization states and an increasing difficulty mechanism