About

1. Brief Introduction

Name: Zhongrenjie(Roger) Lyu
LinkedIn: https://www.linkedin.com/in/zhongrenjie-lyu-5588a928a/
AI Cloud Infra&Backend Development&Distributed System&System Software
Java&Python&Golang&C++
Tecent&Infinigence-AI&CEC&PKU CS&Momenta Intern

2. Education Experience

Peking University (2024.09- 2027.06)
- M.D.，System Software
- Rank:14% GPA:3.7/4.0
University of Electronic Science and Technology of China (2020.09-2024.06)
- B.Eng.，Software Engineering
- Rank:3/181 GPA:3.99/4.00
Chongqing Nankai Secondary School (2017.09-2020.07)

3. Work Experience

Tencent — CSIG (2025.09-2025.10)

Shenzhen

Intern
- Agent & Cloud.
- Multi-Cluster & Multi-Environment Deployment Architecture / Network Traffic Isolation Refactoring
  
  Led the architectural upgrade of multi-environment deployments across Kubernetes clusters in Frankfurt, Singapore, and Guangzhou.
  Implemented fine-grained, personnel-level traffic routing using Traefik Ingress, and refactored the entire Agent Helm deployment to achieve strict environment isolation between development and testing, significantly improving iteration efficiency.
  Designed and implemented a progressive rollout channel that combines Traefik’s weighted routing with label-based traffic matching, enabling safe, incremental releases of new Agent versions.
- Agent API Refactor with Fiber + Prefork & Graceful Shutdown Mechanism
  
  Led the Agent API service refactoring based on the Fiber framework with Prefork mode, introducing dumb-init to ensure proper SIGTERM signal propagation across multiple worker processes, achieving coordinated graceful shutdowns during Kubernetes rolling updates.
  Designed a cross-process dialogue state tracking mechanism on the local file system, featuring atomic writes and timeout-based cleanup to ensure session consistency.
  This mechanism supports active tasks in completing or recovering safely during shutdown, significantly enhancing service reliability and operational robustness within distributed Kubernetes environments.
- Vibe Coding Paradigm for AI-Augmented Development
  
  Established a unified Vibe Coding paradigm across multiple projects, leveraging Roocode and Spec Workflow MCP to formalize collaborative development modes.
  This approach standardized the interaction between developers and AI agents, enhancing productivity, contextual continuity, and team-wide efficiency in AI-assisted coding workflows.
Canlah.ai（2025.06-2025.09）

Silicon Valley (Remote)

Intern
- Agent
- Canlah Calendar Scheduling – Group Meeting Agent Development
  
  Led the end-to-end development of the Group Meeting Agent module within the Canlah Calendar Scheduling system, leveraging LangGraph and FastMCP frameworks to enable intelligent multi-party coordination, conflict resolution, and availability inference.
  The feature has been successfully deployed to production, and the MVP version is now actively used by students from NUS, NTU, and Stanford University, demonstrating real-world adoption and scalability across academic collaboration scenarios.
Peking University Changsha Institute of Computing and Digital Economy（2025.04-2025.07）

Beijing

Intern
- Computing power scheduling, containerization research, container networks.
Infinigence-AI（2024.09-2025.09）

Beijing

Intern
- AI Cloud Infra.
- Tenant-Level Egress Bandwidth Control & Convergence Design
  
  Innovatively designed and implemented a tenant-level egress bandwidth control and convergence system, leveraging custom VXLAN encapsulation, a CRD-based EgressGatewayPolicy, Linux tc (traffic control), and custom routing tables.
  By deploying custom Pods as distributed egress gateways, the system achieves high-availability bandwidth regulation across multi-tenant environments.
  The distributed reconciliation logic ensures robustness and compatibility with multiple CNI plugins, effectively filling a functional gap in existing EgressGateway implementations.
  This solution has been patented and successfully deployed on clusters with thousands of nodes, maintaining stable performance under heavy workloads.
- Dind (Docker-in-Docker) Refactor & Optimization for Development and Inference Services
  
  Led the complete refactor of the Dind service architecture to overcome limitations in the existing development and inference environments.
  Proposed and implemented improvements across four key areas — filesystem interoperability, controllability, configurability, and resource sharing — by redesigning the service initialization flow and resource management pipeline.
  The upgraded system significantly enhances developer experience, runtime flexibility, and resource utilization efficiency, providing a robust foundation for continuous development and inference workloads.
- GPU Resource Reuse and Oversubscription Mechanism
  
  Developed a custom Container Runtime enabling GPU device sharing between containers, resolving exclusivity issues in Dind-based workloads.
  Designed a Pod-level GPU multiplexing mechanism based on ShadowPod and virtual GPU devices, enabling dynamic GPU oversubscription (Steal GPU) through virtualized resource scheduling.
  This approach increases cluster GPU utilization while maintaining workload isolation.
  The system supports multiple GPU vendors, including domestic chips, demonstrating strong portability and extensibility across heterogeneous environments.
- GPU Resource Management and Pod Mapping System
  
  Researched and implemented a GPU–Process–Container–Pod mapping mechanism, integrating NVIDIA NVML with container task management APIs.
  Built a gRPC-based control-data plane communication layer to expose real-time GPU usage and ownership data for scheduling and monitoring.
  This mapping infrastructure provides critical visibility and telemetry support for GPU resource orchestration and intelligent scheduling in large-scale inference clusters.
- Cloud-Agent for Intelligent Cloud Resource Management (Wenwen Xinqiong Project)
  
  Constructed a Retrieval-Augmented Generation (RAG) system leveraging the company’s MASS platform (LLM + Embedding + Rerank) and FAISS vector database to enhance question-answering accuracy and contextual relevance.
  Developed the MCP Tool based on LangGraph and FastMCP, integrating with cloud management APIs to enable dynamic invocation and automated operational workflows.
  This system extends the capabilities of intelligent agents in cloud operations, combining semantic retrieval, agent reasoning, and autonomous tool orchestration into a cohesive AI-driven cloud assistant framework.
CEC（2024.06-2024.08）

Beijing

Research Intern
- LLM research.
- Conducted research on vertical-domain LLMs, focusing on deployment and evaluation.
- Deployed models via Hugging Face and Ollama, and performed comparative benchmarking against general-purpose LLMs in multiple specialized domains.
Peking University (2023.08-2024.05)

Beijing

Research Intern
- High performance middleware, operating system, distributed system.
- In the context of heterogeneous digital space and digital networking, the software and hardware development (Java language and Netty framework/cryptographic signature acceleration card) is used to realize the digital networking switch for heterogeneous data space in order to achieve the purpose of the interaction between digital networking infrastructure (digital networking positioning system, etc.) and IDS and other types of digital space, the north-south interfaces of digital networking, and the clients, so that the digital space, the warehouses, and the clients All access to the digital network, the multi-party digital space and digital networking and digital networking infrastructure to combine, to achieve high performance, high availability, high stability, cross-space, high compatibility of data sending and receiving, data positioning and access, and through the node to complete the update level of the digital networking reference realization using the data in the data language use of the data associated with the relevant information.
Momenta (2023.02-2023.08)

Beijing

Development Intern
- High precision map storage system.
  - Participate in the development process of map spatio-temporal data management system based on multi-version control, Java Springboot framework and juiceFS file system, use Swagger to build visual API documents, build continuous integration and automated test platform based on jenkins.
  - For the map data bias business requirements, I help the downstream departments to create dynamic link libraries, use JNI to enable Java to interact with them, and design relevant interfaces to decouple the business logic so that the storage and bias operations can be separated to realize the business requirements.
  - In order to let the outsourcing business department use the system, I use logback and slf4j framework to realize log record storage, and use Spring Cloud Gateway for routing and forwarding to achieve the purpose of permission restriction, and packaged docker image.
    -Write python scripts and set up timed statistical map system data under Linux system, automate data import into PostgreSQL and use dataease to configure pg data source to achieve data visualization.
  - For map adjacency judgment (changed from rule engine to data-driven), designed to extract osm map feature data, designed neural network structure for binary classification of map adjacency, with accuracy, recall and F1 Score close to 0.95 on both training and test sets.

4. Awards

2025 Outstanding Student of Peking University
2025 Third-Class Scholarship of Peking University
2024 Outstanding University Graduates of Sichuan Province
2022 National Scholarship
2023 National Scholarship
2021 First class scholarship of UESTC
2022 First class scholarship of UESTC
2023 First class scholarship of UESTC
2022 Third Prize of National Finals of C4-Network Technology Challenge
2022 Third Prize in Southwest Region of C4-Network Technology Challenge
2022 The Honorable Mention of MCM

5. Projects

Design and Implementation of a Digital Networking Switch for Heterogeneous Data SpaceDesign and Implementation of a Digital Networking Switch for Heterogeneous Data Space

2023.10 - 2024.05

related organization: Peking Unversity

https://edu.gitee.com/pkudataspace/repos/pkudataspace/iod-switch

data-switch (2).png

Digital networking is a technical solution to interconnect heterogeneous data spaces using standardized and protocolized ideas. Through the digital networking gateway, the heterogeneity of different data spaces is shielded; through the digital networking switch, the DOIP message forwarding across data spaces is realized; through the digital networking router to manage the correspondence between the identifiers in the heterogeneous space and the digital networking switch.
This project is mainly based on the Java language using Netty framework to implement the heterogeneous data space-oriented digital networking switch, through the use of DOIP-SDK, IRP-SDK, Caffiene high-performance cache components, optimization of caching, optimization of connection pooling based on FixedChannelPool, based on the development of JCE interfaces to use the Sansec encryption card and other designs. Implemented a digital networking switch, through Grafana and built-in Prometheus on the digital networking switch statistics, through the relevant Controller class and Javascript interface call code and its deploy method deployed in a number of cloud servers, can be directly through the DOBrowser switch related to the interface call. In terms of functionality, the Digital Networking Switch can correctly interact with other Digital Networking Switches or Digital Networking Gateways to complete the correct parsing of DOIP protocol requests and the correct forwarding of packets for packet switching; in terms of performance optimization, it optimizes the client connection, caching and other aspects. Meanwhile, this thesis innovatively constructs a cache load test set based on Pareto distribution and conducts targeted tests on performance optimization aspects, and the experimental results show that the Digital Networking Switch achieves high-performance, high-availability, high-stability, cross-space, and high-compatibility data sending and receiving, data positioning, and acquisition.

Children’s Home Intelligent Safety Detection System

2021.09 - 2023.01

related organization: University of Electronic Science and Technology of China

Use python language crawler to crawl the dataset and write code for data enhancement to expand the dataset, use labelImg to label the dataset, and train the model on the crawled dataset based on pytorch, the open source project YOLOv5 and OpenPose. Self-modify the YOLO source code and write the rule engine to realize the determination of children’s dangerous actions and dangerous coefficients, realize real-time warning, modify the loss function and fuse the two models to improve the detection accuracy.
Write test cases and test plans, use pytest framework for interface testing, and write related documents. The project was selected as the outstanding student work of the class of 2020 in the School of Information and Software Engineering of the University of Electronic Science and Technology for exhibition, and was shortlisted for the national competition of the Challenge Cup and Computer Design Competition, and won the third prize of the finals.

Monitora-Intelligent Network Inspection System Based on P4 and SRv6

2022.04-2022.09

related organization: University of Electronic Science and Technology of China

In this project, a P4-based network monitoring system is designed to realize real-time and efficient monitoring of the network. Through the P4-based active telemetry system, the appropriate number and special format of probe data packets are used to replace the normal data packets for network telemetry, which reduces the telemetry overhead and improves the payload ratio of the data packets. Efficient real-time network telemetry is realized.
Based on the simple and flexible routing control capability of segment routing (SR, segment routing) to dynamically specify the probe path by changing the SR label and its ordering at runtime. We can add SR label stacks to the probes to support getting a comprehensive view of the network. Also, by using a ring probe path, a single probe point can serve as both the transmitter and receiver of the probe, which reduces the complexity of synchronizing and coordinating probes among multiple probe points. On the other hand, in order to reduce redundant telemetry data, a telemetry data indication field is added to the probe data group to specify the telemetry data to be collected. Finally, embedding the state information inside the programmable device into the probes can be achieved by customizing the data grouping processing logic through the customization capability of the programmable device.