Googler & Berkeley Alumni
Experience
2022 — Now
San Francisco Bay Area
Ads Recommender System
Data Infra
2021 — 2022
Redmond, Washington, United States
Capstone Project:
Tracking System: Architected and implemented Microsoft cloud deployment tracking and management system with SpringBoot, used Hibernate to persist data in MySQL database. Implemented Redis Cluster as cache database, boosted QPS limit by 82%. Automated CI/CD with Jenkins. Reduced response time by 55% with estimated saving of $15.27 million/year.
Analytics Service: Developed offline analytics service with Hive and Spark. Used ElasticSearch as search engine for delivery and work order log retrieving. Built interactive frontend with React and ECharts to visualize current status and progress.
Real-time ETA Service: Collected real-time orders with Flume, sent to Kafka message queue, performed real-time feature computing with Storm and stored data in HDFS, powering downstream data-driven ETA Service. Reduced ETA error by 78%.
Microservices Governance: Deployed Consul for service discovery, with Raft Algorithm to ensure system consistency, and Gateway as rate limiter. Followed MDM and CQRS design pattern to build Main Data Platform, achieved 99.95% availability.
Cloud Service Deployment: Containerized services with Docker, deployed to Kubernetes cluster on Azure.
2020 — 2021
Beijing, China
Autonomous Car Planning Module: Pioneered in designing, developing and deploying onboard robust, scalable and low-latency obstacle decision service using C++, Bazel, Protobuf and Docker. Raised task completion rate by 42%. Responds to 1,000,000,000+ daily requests. Generated $4.58 million/year extra revenue comparing with previous module.
Fleet Analytics Service: Used gRPC protocol for microservices communication. Built the distributed system with SpringCloud, Zookeeper, MySQL, Hive and Spark. Provided API to 60+ third-party services. Responds to 50,000+ daily queries.
Fleet Scheduling Service: Developed the backend with Golang, used Go Micro as microservices framework, GORM for Object-Relational Mapping and Redis, MySQL for database. Reduced latency by 43%.
Reinforement Learning: Proposed a Hierarchical Multi-Agent Reinforcement Learning Algorithm to handle trajectory prediction problem in dense traffic with complicated agent interactions. Raised task completion rate to 94%, surpassing all previous work.
AI-based Planning Algorithm: Trained XGBoost, Deep Learning, designed HGAT (Hierarchical Graph Attention Network) models for overtake decision making with PyTorch. Reduced FDE by 82%, FLOPs by 77% in deployment.
Chengdu, Sichuan, China
Education
2021 — 2022
University of California, Berkeley
Master’s Degree
2021 — 2022
2016 — 2020
Sichuan University
Bachelor’s Degree
2016 — 2020