I am a Ph.D. student in the Department of Computer Science and Engineering at the Hong Kong University of Science and Technology (HKUST), advised by Prof. Binhang Yuan and Prof. Wentao Zhang. My research focuses on the systems layer between heterogeneous raw data and large language model training: making data preparation easier to compose, distribute, reproduce, and operate with Ray and Apache Spark.
I am currently a Qingyun Program Research Intern with Tencent Hunyuan’s Foundation Model Department in Beijing, where I work on data infrastructure for large-scale LLM data preparation, including operator management and distributed pipelines for deduplication, clustering, and document and video processing.
Before starting my Ph.D., I was a Research Assistant at Peking University from January 2025 to January 2026, where I worked with Prof. Wentao Zhang on LLM data preparation systems, including DataFlow. Earlier, I was a full-time Research Assistant at Sichuan University from July 2023 to June 2024, working with Prof. Jizhe Zhou on image manipulation detection and localization. This progression from reproducible vision benchmarks to general-purpose data preparation systems motivates my current work on scalable and reusable data workflows.
I received my B.Eng. in Computer Science and Technology from the College of Computer Science at Sichuan University in June 2023. My undergraduate thesis was recognized by the university as an Outstanding Undergraduate Thesis.
Research impactGoogle Scholar787citationsh-index10i10-index11
Project lead · Accepted at SIGMOD'27LLM Data InfraRay
Project
A composable data-pipeline and workflow system for LLM data preparation, spanning data generation, cleaning, evaluation, and orchestration.
My contribution
Project lead and core contributor, coordinating a team of more than 10 students across system design, implementation, code review, integration, and releases. My engineering contributions span pipeline compilation and execution, operator and registry design, storage, serving, CLI tooling, and RayOrch integration.
A modular training and evaluation codebase and comprehensive benchmark for image manipulation detection and localization.
My contribution
Joint first author, codebase designer, and coding lead. I designed the modular training and evaluation framework and have maintained its metrics, model zoo, CLI, packaging, releases, and tests.
A Ray-powered acceleration layer that turns MinerU PDF-to-Markdown parsing into a scalable multi-GPU and multi-node data pipeline.
My contribution
Project initiator and core developer. I built the initial package and release workflow, then refactored the sequential stages into an asynchronous pipeline with cross-stage overlap, while maintaining benchmarks and compatibility.
A programming model and distributed execution engine for foundation-model data preparation, preserving parent-child lineage, ordering, completion, and result assembly as each input expands into multiple records.
My contribution
First author and project initiator. I designed the dataflow model and runtime for ordered variable-cardinality expansion, compiler validation, and parent-scoped failure handling, and integrated the Ray-based executor into DataFlow.
