Virginia Tech

College of Engineering
PhD candidate in Computer Science

University of North Carolina at Chapel Hill

Gillings School of Public Health
Master of Science in Biostatistics

Zhejiang University

School of Mathematical Sciences
Bachelor of Science in Statistics
&&
School of International Studies
Bachelor of Arts in English Literature and Linguistics

Graduate Research Assistant

Computer Science, Virginia Tech, Blacksburg, USA

• • Applied rigorous statistical methods and core computer science principles to investigate large language model (LLM) behavior, bridging the gap between theoretical guarantees and real-world performance in open-world settings.

Graduate Research Assistant

Biostatistics, UNC-Chapel Hill, Chapel Hill, USA

• • Architected scalable AI/ML frameworks that unify heterogeneous biomedical data (single-cell & bulk RNA-seq, EHRs) to generate precise patient-level insights and streamline large-scale comparative analyses.

Student Researcher

Statistics, Zhejiang University, Hangzhou, China

• • Applied advanced ML/DL techniques to both quantitative finance and computational biology, delivering state-of-the-art market-forecasting/trading systems and precision RNA-seq & biomarker pipelines that markedly boost predictive accuracy and operational efficiency.

Software Design Engineer Intern

Alibaba, Hangzhou, China

• • Replaced a legacy static-path system with a Java-based Dynamic Source Routing (DSR) algorithm, embedding a node-energy score that raised end-to-end packet delivery from 88 % to 97 % and cut average latency by 30 % during 500-node stress tests under highly dynamic traffic.

Quantitative Research Intern

Tenbagger Capital Management, Hangzhou, China

• • Built and maintained a web-scraped financial warehouse (5 M+ records across 1,000 equities) capturing key metrics—transaction rate, gross-profit margin, leverage ratio, etc.—to enable rapid factor research.
• • Developed an automated Python long-short strategy driven by those factors, delivering ≈15 % annualized alpha and Sharpe 1.3 in 3-year walk-forward tests.

LensLLM: Unveiling Fine-Tuning Dynamics for LLM Selection

VirginiaTech (Advisor: prof.Dawei Zhou)

• • Formulated a PAC-Bayesian bound that isolates the pre-power vs. power phases in LLM fine-tuning and pinpoints the phase- transition threshold.
• • Designed LENSLLM, an NTK-based rectified scaling law that predicts a model's full-data performance after observing only a small training subset.
• • Built a progressive sampling algorithm that cuts fine-tuning compute by up to 88.5 % on Pytorch.
• • Fine-tuned on 7 LLM families (OPT, T5, GPT-2, mT5, BART, etc.), achieving 91.1 % relative accuracy and 85.8 % Pearson correlation, outperforming five baselines across FLAN, WikiText, and Gigaword tasks.

Cell-type Specific RNA-sequence Genetic Deconvolution

UNC-Chapel Hill (Advisor: prof. Yun Li)

• • Spearheaded a scalable Bayesian Gaussian–based probabilistic framework that integrates single-cell and bulk RNA-seq data through a nonlinear dynamic covariate system, producing precise sample-level cell-type expression estimates.
• • Delivered up to 123.1 % higher Pearson correlation than leading methods across multiple datasets, evidencing superior accuracy and robustness.

Mixed Effects Modeling via Deep Learning

UNC-Chapel Hill (Advisor: prof. Baiming Zou)

• • Engineered a deep-learning semiparametric regression pipeline in R that extracts high-impact biomarkers from EHRs while adjusting for complex, nonlinear confounding.
• • Built Linux-shell workflows that scale comparative-effectiveness analyses from simulated to real-world datasets, sharply reducing preprocessing time and ensuring reproducibility.

Development of Stock Investment System Based on Big Data Analysis and Statistical Optimization

Zhejiang University (Advisor: prof. Peng Zhang)

• • Built a Python/TensorFlow mid-frequency stock-forecasting system (RSRS → GBDT → 4-state Gaussian HMM) on a 10-year, 800+ equity dataset, boosting hit-rate from 52 % to 67 % with α = 8.4 %, Sharpe = 1.6, and capped drawdown at -5.8 %.
• • Deployed models on a two-node HA Linux cluster via a low-latency trading platform (MATLAB UI, Java backend), achieving <100 ms execution for 1k+ daily orders and cutting reconciliation time by 85 %.