A junior student major in Computer Science with Artificial Intelligence.
I am broadly interested in generative models, AI agents, WAM, and VLA.
My background is in deep learning research and engineering, with experience in foundation model development, model adaptation and post-training, research implementation, benchmarking, and ML systems engineering.
＿＿Research at NC State University – NiceLab＿＿＿＿＿＿＿＿＿Jul - Aug 2026
I designed two quantum integration architectures—late quantum bottleneck and quantum self-attention—and validated their performance on drone-signal spectrograms and FER2013. Findings reveal that embedding quantum self-attention into pairwise feature interactions yields consistent improvements over Vision Transformer baselines, especially under low SNR conditions, while bottleneck placement only offers local advantages.
Team Leader | Mentor: Prof. Meng Ming＿＿＿＿＿＿＿＿＿2025.12 - Now
Tech: Python, PyTorch, LoRA, Diffusion Models, Gradio, CUDA, GPU Deployment
Designed and implemented a dual-stage LoRA fine-tuning strategy to improve the Wan2.2-Animate 14B model's performance on 2D character animation and long-range motion generation. I deployed the full 14B model locally (PyTorch 2.2, CUDA 12, FlashAttention2, DeepSpeed-ZeRO) and developed two coordinated LoRA modules: a low-noise LoRA to stabilize stylized 2D facial structures and suppress human-face prior collapse, and a high-noise LoRA to enhance distant-view pose consistency and temporal coherence.
Additionally, I built a complete Gradio + FastAPI platform supporting inference, LoRA switching, visualization, and a one-click fine-tuning pipeline (“LoRA Furnace”). This system delivered measurable improvements in 2D style fidelity and long-range motion stability, offering a scalable framework for multi-style, multi-scale animation model optimization.
＿＿＿＿＿Personal Competition＿＿＿＿＿＿＿＿＿2025.12 - 2026.01
Developed a time-series machine learning pipeline to forecast MDS-UPDRS motor and non-motor scores for Parkinson's disease patients using longitudinal peptide and protein profiles provided by the AMP®-PD Knowledge Platform. Implemented models under strict non-leakage constraints enforced by the competition's Python API, generating visit-level predictions at 0, 6, 12, and 24-month horizons.
Experimented with sequence-aware regression architectures and feature representations for sparse, irregular multi-omics measurements; evaluated performance via SMAPE. Through this work, gained practical experience handling high-dimensional biological time-series data, domain shift between Parkinson's and control cohorts, and the challenges of biomarker discovery in neurodegenerative diseases.
＿Team Leader | Mentor: Prof. Zhang Qin＿＿＿＿＿＿＿2024.09 - 2025.01
This project is a smart home control system using the MQTT protocol. It integrates data publishing, real-time monitoring, historical data management, and a graphical user interface (GUI), ideal for teaching demos, prototyping, and IoT coursework.
＿＿Team Leader | Mentor: Prof. Zhang Qin＿＿＿＿＿＿＿＿＿2023.09 - 2024.06
Proposed an improved TFBSS algorithm to address the challenge of separating mixed source signals containing Gaussian components. The method integrates time–frequency representation (TFR) with independent component analysis (ICA), optimizing both the TFR matrix and the joint diagonalization process to achieve efficient signal separation. Extensive simulations demonstrated that the enhanced TFBSS algorithm provides higher accuracy and robustness in separating source signals with Gaussian characteristics, offering valuable insights and potential applications for the field.
点击查看
＿＿Team Leader | Mentor: Prof. Zhang Qin＿＿＿＿＿＿＿＿＿2024.09 - 2025.01
Explored the integration of gesture recognition with large-scale models and optimized convolutional neural networks (CNNs) for robust feature extraction. By incorporating multi-view image data, developed a novel network architecture capable of accurately capturing hand features and estimating gesture poses. My contributions significantly improved recognition accuracy and real-time performance, highlighting strong application potential in smart devices, virtual reality (VR), and human–computer interaction.
(Detailed project write-ups available at: qushihan.space/about.html)
Gesture