Hi, I am Banruo.
劉般若
MCS student at UIUC building systems for agentic AI, LLM serving, compound AI applications, and distributed infrastructure.
About
I am a Master of Computer Science student at the University of Illinois Urbana-Champaign, advised by Prof. Fan Lai, and expect to graduate in December 2026. I transitioned from the Ph.D. program to the MCS program after two years of graduate research. My work focuses on the systems side of AI: serving and scheduling for agentic and compound AI applications, GPU infrastructure, and distributed systems. I received my B.Eng. in Computer Science from Tsinghua University.
Education
University of Illinois Urbana-Champaign
Advisor: Prof. Fan Lai. Transitioned from the Ph.D. program to the MCS program.
Tsinghua University
Advisor: Prof. Youyou Lu.
Experience
Microsoft Azure Research
Characterized production-scale GitHub Copilot agentic coding workloads and identified system optimization opportunities for coding agents and their harness.
Google System Research
Led Magnet, a vLLM-compatible serving engine for real-time agentic LLM applications represented as multi-stage dataflow graphs.
Publications
Agentic Coding in the Wild: Characterizing GitHub Copilot Traces at Production Scale
Preprint. Banruo Liu, Haoran Qiu, Íñigo Goiri, Rodrigo Fonseca, Ricardo Bianchini, Esha Choukse.
Production-scale characterization of agentic coding workloads, revealing key system optimization opportunities for agent serving.
Magnet: Real-Time Agentic Serving via Just-in-Time Dataflow Execution
In submission. Banruo Liu*, Tony Hong*, Yeounoh Chung, Brent Stephens, Arvind Krishnamurthy, Fan Lai.
Serving engine for real-time agentic LLM dataflows with just-in-time execution and topology-aware GPU allocation.
Single-agent or Multi-agent Systems? Why Not Both?
In submission. Mingyan Gao*, Yanzi Li*, Banruo Liu*, Yifan Yu, Phillip Wang, Ching-Yu Lin, Fan Lai.
Empirical study of single-agent and multi-agent LLM systems; proposes routing and cascade strategies for cost-quality tradeoffs in agentic applications.
Compass: SLO-aware Query Planner for Compound AI Serving at Scale
VLDB 2026. Banruo Liu, Wei-Yu Lin, Minghao Fang, Yihan Jiang, Fan Lai.
SLO-aware query planner for compound AI pipelines; improves service goodput by 2.4-5.1x and reduces deployment cost by 3.8-4.5x.
JITServe: SLO-aware LLM Serving with Imprecise Request Information
NSDI 2026. Wei Zhang*, Zhiyu Wu*, Yi Mu, Rui Ning, Banruo Liu, Nikhil Sarda, Myungjin Lee, Fan Lai.
LLM serving scheduler for latency-sensitive, deadline-sensitive, and compound requests under response-length and dependency uncertainty.
High-level Programming for Application Networks
NSDI 2025. Xiangfeng Zhu, Yuyao Wang, Banruo Liu, Yongtong Wu, Nikola Bojanic, Jingrong Chen, Gilbert Bernstein, Arvind Krishnamurthy, Sam Kumar, Ratul Mahajan, Danyang Zhuo.
High-level language, compiler, and controller for programmable application networks in microservice systems.
Towards a Flexible and High-Fidelity Approach to Distributed DNN Training Emulation
APSys 2024. Banruo Liu, Mubarak Adetunji Ojewale, Yuhan Ding, Marco Canini.
Distributed DNN training emulator that executes real training nodes while emulating networked collective communication.
Application Defined Networks
HotNets 2023. Xiangfeng Zhu, Weixin Deng, Banruo Liu, Jingrong Chen, Yongji Wu, Thomas Anderson, Arvind Krishnamurthy, Ratul Mahajan, Danyang Zhuo.
Position paper on application-defined networking for programmable microservice communication and application-aware network functions.
* denotes equal contribution where applicable.