Xiaoquan KongAI Engineer & Researcher I build AI systems that work at scale—from research prototypes to production platforms serving millions. My work focuses on agentic AI, retrieval-augmented generation, and reinforcement learning. At Baidu and Alibaba, I built conversational AI systems deployed across millions of vehicles serving large-scale production workloads. I hold a MEng from Duke University and am a Google Developer Expert in Machine Learning. Email: u1mail2me@gmail.com / GitHub / LinkedIn / Google Scholar |
|
Research Interests
- Agentic AI Systems: Multi-agent architectures, tool use, agent coordination
- Retrieval-Augmented Generation: Multi-modal RAG, evidence-based reasoning
- Reinforcement Learning: RLHF for LLM alignment, offline RL, model-based RL
- Production ML Systems: Scalable deployment, low-latency inference, system optimization
Research & Development
Reinforcement Learning Curriculum Development [course site] [course]
Duke University
Designed curriculum for graduate RL course (AIPI 590) covering tabular methods to modern RLHF. Built complete alignment pipeline from scratch—reward modeling, PPO training, safety evaluation—demonstrating significant improvements in model safety. Students implement core algorithms from first principles, learning to align language models with human preferences.
Educational AI Development
Duke CREATE (Center for Research & Engineering of AI Technology in Education)
Developing QUBIT, an AI programming assistant that teaches through explanation rather than direct code generation. Implementing scaffolded prompting and constraint-based hints to guide students through problem-solving while maintaining engagement.
Selected Projects
A minimal implementation (~300 lines) exposing the complete control flow of agentic systems. By avoiding framework abstractions, it reveals how tool-calling and multi-agent coordination actually work—enabling developers to understand and modify agent behaviors at the implementation level. Supports major LLMs via LiteLLM.
PyTorch-compatible ML framework built from scratch—autodiff engine, neural networks, optimizers. Designed for understanding deep learning fundamentals with drop-in replacement capability.
Enterprise-grade RAG processing text, images, videos, and audio with evidence-based answer generation. Built with Vertex AI, Pinecone, and Gemini-1.5-Pro for auditable decision-making.
Suite of Chinese NLP tools including spaCy models, character decomposition library, and educational tokenizer. Widely adopted in industry and academia with strong community engagement.
Books
Xiaoquan Kong, Guan Wang • Packt Publishing, 2021 • [Amazon]
Production-grade chatbot development. Foreword by Alan Nichol (Rasa Co-founder & CTO).
Guan Wang, Xiaoquan Kong • BroadView Press, 2024 • [JD.com]
Traditional NLP to modern LLM applications. Foreword by Ines Montani (spaCy Co-founder).
Xiaoquan Kong • Online Book • [read online]
Master AI agents by building a production-ready framework from scratch. Covers agent architecture, multi-agent systems, and MCP integration.
Xiaoquan Kong • Online Book • [read online]
Build a PyTorch-compatible framework from scratch—autodiff, optimizers, modules—culminating in a working GPT implementation.
Selected Publications
550+ citations, h-index: 7 • [Google Scholar]
Ren-Yan Duan*, Xiaoquan Kong*, et al. • PloS ONE, 2014
Min-Yi Huang*, Xiaoquan Kong*, et al. • PeerJ, 2016
* denotes equal contribution
Honors & Awards
- Google Developer Expert in Machine Learning, 2018 - Present
- TensorFlow Addons Codeowner, Led development of CRF module for NLP
- Baidu Chief AI Architect, First cohort, 2019
- Google Summer of Code Mentor, TensorFlow team, 2022
Community & Open Source
- GitHub: Active contributor with significant community engagement across multiple repositories
- Open Source: Key contributor to TensorFlow, spaCy, Rasa, Google Agent Kit
- Speaking: Regular speaker at technical conferences and workshops on AI/ML topics
Education
Duke University