Query Optimization
Robust Selinger-style optimization, uncertainty-aware plan construction, and PostgreSQL kernel work.
Research
Building reliable systems for data, retrieval, and AI evaluation.
My current interests center on query optimization and processing, vectorized databases and RAG, and data center networks.
Database-Centric RAG
PostgreSQL, pgvector, Lucene, Kafka, and Spark pipelines for retrieval and analytics.
Multimodal AI Systems
VLM-as-a-judge systems for large-scale safety and policy compliance evaluation.
Education
Academic Background
Aug 2024 - May 2026
Duke University
Master of Science in Computer Science, GPA: 3.81/4.0
Durham, NC. Advisors: Prof. Jun Yang and Prof. Xiaowei Yang.
Sep 2020 - Jun 2024
Central University of Finance and Economics
Bachelor of Data Science, Guozhi Hsu Talent Program in Data Science, GPA: 3.73/4.0
Beijing, China.
Projects
Selected Work
Research and engineering projects across database kernels, retrieval infrastructure, streaming data, and vision-language model evaluation.
A Practical Flexible Selinger-Style Robust Query Optimizer
- Proposed a robustness-inside-Selinger optimizer integrating selectivity uncertainty during DP-based plan construction.
- Enabled runtime robust plan identification in PostgreSQL V16.2 with pluggable objectives and plan diversification strategies.
A Database-Centric RAG System Leveraging LLMs
- Developed a full-stack RAG system with PostgreSQL, pgvector, and Qwen3 for record retrieval by names, keywords, and descriptions.
- Built real-time analytics with Spark Streaming and Kafka, plus a Lucene-based multi-field fuzzy search engine.
Content Moderation in Web and Networked Systems
- Developed a multimodal VLM-as-a-judge system for online content moderation across text, image, and video.
- Built data acquisition, preprocessing, prompt-based inference, and evaluation pipelines for models including Qwen, DeepSeek, and LLaMA.
Cross-Modal Hard Feature Mining for Few-Shot Learning
- Proposed DeltaAug, a CLIP-based data augmentation strategy inspired by hard-class behavior.
- Achieved 76.16% accuracy on 16-shot classification over 11 datasets, outperforming the prior Wise-FT baseline.
Publications
A Practical Query Optimization System with Built-in Robustness
Haibo Xiu, Xuan Chen, Rahul Raychaudhury, Pankaj K. Agarwal, Jun Yang. Under submission, VLDB 2026 expected.
Delta-Aug: Cross-Modal Hard Feature Mining for Few-Shot Learning
Xuan Chen. Published in the 2024 6th Asia Conference on Machine Learning and Computing.
Skills
C/C++
Java
Linux
PostgreSQL Kernel
MySQL
Python
PyTorch
LLMs
VLMs
Docker
TypeScript
Vue 3
Honors
Recognition
6th "Hello World" Algorithm and Programming Competition
Special Prize, Team Award, 1st - 2023
Mathematics Competition for Chinese College Students
Second Prize - 2022
Academic Research Scholarship, School of MSE, CUFE
First Prize - 2021
Contact
Available for research conversations and collaboration.
Based in Durham, North Carolina.
- Email xc241@duke.edu
- Phone 919-358-2845
- Homepage cx0222.cn
- Location Durham, NC