B.S. and M.S. from Wuhan University. Tech Lead of Large Language Models at PayPal AI Platform, powering thousands of models and tens of billions of inference requests. Focusing on large language models, reinforcement learning, Graph, Post-Training, agents, and LLM quantization & inference.

10+ national invention patents, multiple academic papers published. Previously worked at Baidu on high-performance data science engines and AI/ML platforms.

Selected work

• Adaptive-Boundary-Clipping GRPO: Ensuring Bounded Ratios for Stable and Generalizable Training
  Chi Liu, X Chen | 2026
  paper | github

• Rethinking GSPO: The Perplexity-Entropy Equivalence
  Chi Liu | 2025 | Cited by 1
  paper

• Fraud Detection Through Large-Scale Graph Clustering with Heterogeneous Link Transformation
  Chi Liu | 2025 | Cited by 1
  paper

Patent

• Container Image Processing Method and Apparatus, and Non-Transitory Computer-Readable Storage Medium
  Chi Liu, Ziheng Li, Kai Chen, Xiaoning Yu, Hui Han
  US Patent US11880341B2, 2024
  Assignee: Beijing Baidu Netcom Science and Technology Co Ltd
  patent

• A Method, Device, Equipment and Medium for Realizing Joint Modeling
  Hui Han, Kai Chen, Chi Liu, Jiayi Yang
  Chinese Patent CN112182635B, 2024
  Assignee: Beijing Baidu Netcom Science and Technology Co Ltd
  patent

• Operating Environment Acquisition Method, Device and Electronic Device
  Kai Chen, Hui Han, Xiaoning Yu, Chi Liu, Ziheng Li, Jiayi Yang
  Chinese Patent CN110908675B, 2023
  Assignee: Beijing Baidu Netcom Science and Technology Co Ltd
  patent

Open Source Contributions

Active contributor to key open-source projects in the LLM ecosystem, including a ~18.7% performance improvement for the vLLM scheduler and a ~6.5× speedup for Verl via GPU device placement fix. Contributions span LLM inference engines (vLLM, SGLang), RL training frameworks (Hugging Face TRL, Verl), MCP frameworks (FastMCP), and agentic frameworks (CrewAI). View full contribution list →

Public talks

• From MLOps to LLMOps: Building an AI for Data Platform Supporting Thousands of Models and Tens of Billions of Inference Requests (QCon 2024)
  Sharing how PayPal built a unified end-to-end enterprise AI platform to break down data silos and support thousands of models with tens of billions of inference requests; and practical experience scaling from MLOps to LLMOps, covering LLM inference optimization, RAG & Multi-Agent low-code frameworks, semantic caching, and hallucination detection.

(last updated: Mar 2026)