Ph.D. student focused on efficient AI systems, long-context inference, and intelligent security.
Efficient AI Systems | Academic Portfolio
Designing efficient computation for the long-context era of intelligent systems.
I am a Ph.D. student at the National University of Defense Technology, working at the intersection of long-sequence large language model inference, attention optimization, and intelligent security. My research focuses on turning advanced model capability into deployable, resource-aware systems. In September 2026, I will join Professor Trevor E. Carlson's group at the National University of Singapore as a visiting Ph.D. researcher, where I will continue exploring efficient AI systems, LLM serving, and hardware-aware inference optimization.
Research Index
2026
Core Directions
- Long-context LLM inference
- Efficient attention design
- Dynamic memory scheduling
- Neural web attack detection
SVD Compression
Routing
Security
Acceleration
Education
Academic training with a strong foundation in systems, software, and intelligence.
My academic path moves from software engineering fundamentals to advanced research on efficient AI computation, with a continued focus on rigorous engineering and real-world deployment.
2023 - Present
National University of Defense Technology
Ph.D. Student in Large Language Model Acceleration
College of Computer, Key Laboratory of Advanced Microprocessor Chips and Systems, Changsha, China.
2019 - 2023
Xiamen University
B.S. in Software Engineering
Comprehensive GPA: 3.72/4.00, Rank: 2/56. Built a strong background in AI, architecture, networks, databases, and software systems.
Publications
Selected publications across efficient AI systems, accelerator design, and intelligent security.
GRACE-Infer: A Stage-Aware Multi-Granularity Framework with Metadata Reuse for Transformer Inference
Stage-aware multi-granularity transformer inference framework with metadata reuse, now positioned toward CCF-A computer architecture venues.
TRIDENT: An End-to-End Streaming Accelerator for TriSpGEMM
End-to-end streaming accelerator research for TriSpGEMM, now published at DAC and positioned within high-performance systems and accelerator design.
E-WebGuard: Enhanced Neural Architectures for Precision Web Attack Detection
Enhanced neural architectures for precise web attack detection, improving real-world classification performance in security tasks.
Projects
Research projects shaped by systems optimization and measurable outcomes.
LLM Frameworks
MGBH for Long-Sequence Inference
Developed a multi-granularity bit-level hierarchy with dynamic memory management for faster long-context execution.
- 1.45x latency reduction over dense attention
- 2.89x over sparse-band baselines
Attention Optimization
Dual-Path Attention Pipeline
Studied SVD compression, dynamic head routing, and sliding-window strategies for efficient large-model serving.
- Reduced memory pressure in long sequences
- Supported publishable algorithmic results
AI Security
Integrated Web Attack Detection Models
Combined traditional machine learning with deep neural models for precise web application attack detection.
- Char-SVM, CNN-SVM, CNN-Bi-LSTM
- 99.60% binary classification accuracy
Research Experience
A research trajectory spanning algorithm design, software frameworks, and practical security systems.
Jun. 2025 - Dec. 2025
Software Framework Optimization for Long-Sequence LLM Inference
Developed the Multi-Granularity Bit-Level Hierarchical framework by integrating bit-level aggregation and dynamic memory management for long-sequence inference.
May 2024 - May 2025
Algorithmic Optimization for Long-Sequence LLM Inference
Explored SVD compression, dynamic head routing, and sliding-window methods to optimize memory use and computation in large language model inference.
Jun. 2022 - Sep. 2023
Enhancing Web Attack Detection with Neural Networks
Designed hybrid learning pipelines for web attack detection and outperformed earlier neural baselines on binary classification tasks.
Contact
Open to research collaboration, academic visibility, and future-facing system design.
This site is structured as a deployable academic portfolio and can continue evolving into a full research homepage with papers, project notes, and application materials.