Projects

Systems-oriented projects with measurable gains in performance, efficiency, and detection accuracy.

These projects highlight a research style that moves from algorithmic insight to implementation, optimization, and quantified outcomes.

2025 NUDT Framework Optimization

Software Framework Optimization for Long-Sequence LLM Inference

Developed the Multi-Granularity Bit-Level Hierarchical framework to optimize long-sequence LLM inference through bit-level aggregation and dynamic memory management.

1.45x latency reduction over dense attention
2.89x speedup over sparse-band baselines
Minimal accuracy loss in BERT-based experiments

2024 - 2025 NUDT Algorithm Design

Algorithmic Optimization for Long-Sequence LLM Inference

Investigated scalable methods for inference acceleration in long-context settings, including SVD compression, dynamic head routing, and sliding-window execution strategies.

Targeted memory and compute bottlenecks in long-context workloads
Developed the core ideas behind GRACE-Infer
Established a reusable optimization pipeline for future work

2022 - 2023 Xiamen University AI Security

Enhancing Web Attack Detection with Neural Networks

Designed a family of hybrid models for web attack detection using Char-SVM, Char-LSTM, CNN-SVM, and CNN-Bi-LSTM architectures.

Reached 99.60% binary classification accuracy with Char-SVM
Outperformed earlier CNN-LSTM style baselines
Connected machine learning and deep learning pipelines in one system

2020 Tencent Internship

Data Platform Optimization at Tencent PCG Video Technology

Optimized an internal data platform with Hadoop and Spark, improving data synchronization and processing efficiency while contributing to data privacy protection work.

31% improvement in data synchronization
8% increase in processing efficiency
Contributed to a privacy-related patent filing