Apr 2021 – Dec 2024 · 3 yrs 8 mos · Bellevue, Washington, United States

• Computer Vision Ergonomic Risk Detection System (Horus)
• Designed and led implementation across two engineer teams to build a computer vision system (Java-based) called Horus. Horus aims to reduce Amazon station associates’ injuries due to high-risk movements. Horus detects high-risk movements from recorded videos in stations. Once a risk is detected, Horus sends a coaching notification to corresponding associate, which contains video highlighting associate’s high-risk movements and personalized recommendation.
• Implemented RESTful APIs using AWS API Gateway, AWS Lambda and AWS S3 for handling video upload/download with authentication and rate limiting.
• Built video clipping component using AWS Lambda and AWS MediaConvert, which enables associates to view 5-second video clips of high-risk movements rather than the whole video. It reduced coaching view time by 91%.
• The system reduced ergonomic risk score by 18% in North America stations.
• Self-Service Data Aggregation System
• Designed and implemented a SQL-based data aggregation system with a React frontend webapp, replacing existing system requiring to onboard data streams by AWS CDK and define data aggregation by complicated Java code. The new system simplifies process by supporting simple JSON template to onboard data streams and defining aggregation logic by SQL.
• Implemented data ingestion layer using AWS SQS and Lambda and data aggregation layer using EventBridge and Lambda. Integrating with database AWS Timestream for conducting data aggregation on large-scale data with low latency.
• Utilized AWS AppConfig to manage user-defined templates for data normalization (Apache FreeMarker) and data aggregation (SQL), leveraging incremental rollout and auto-rollback for new changes to reduce blast radius
• The system ingests/processes 331,000 events per hour and reduces new data aggregator onboarding time by 70%

Jul 2019 – Mar 2021 · 1 yr 8 mos · Bellevue, Washington, United States

• [Last Mile] MBE(Manage By Exception) Tech - the team focuses on helping Amazon make smooth deliveries by building systems to detect anomalies from millions of Amazon operation events and assign corresponding action items across qualified resolvers to maximize operator utilization.
• New Retrigger System of Anomaly Detection Engine
• Retrigger is a deferred evaluation mechanism of anomaly detection engine, e.g., users can configure an anomaly detection rule such that for each incoming order update, an timer is set up and evaluate the order status at the configured time point(e.g., 90 minutes prior to promised delivery time ) to determine if there is any anomaly
• Designed and implemented a new Retrigger system by using AWS Lambda, DynamoDB and CloudWatch Event Rule
• By replacing old Retrigger system based on AWS Step Functions, the new system reduced the cost by an estimated $110K(9.6% of full-year budget) in 2020 and remove the scalability bottleneck of the old solution as AWS Step Functions has a hard limit of 1 million open executions
• Mission Control
• Mission Control distributes tasks across qualified resolvers to maximize CO operator utilization through expertise-based assignment
• Utilized AWS Lambda, SQS, and Connect to implement an event-driven assignment engine. It assigns tasks based on task priority and CO operators’ bandwidth. Integrating with AWS DynamoDB for managing task lifecycle.
• Used AWS SQS to achieve failure isolation and rate limiting for assignment engine.
• Implemented a CI/CD pipeline with AWS CDK for automatic build and deployment. Leveraged AWS CodeDeploy to configure Blue Green Deployment to achieve incremental rollout and auto rollback.
• SLA(Service-level agreement) Breaching Detection System
• Designed and implemented SLA Breaching Detection System based on AWS Lambda and DynamoDB, which enables users to configure SLA rules such that notification/action items will be automatically triggered once a SLA is breached

May 2018 – Sep 2018 · 4 mos · Hartford, Connecticut Area

• Used Spring, SpringMVC to develop a Java web application which provides multiple service about genomics data, such as Data Visualization, Outlier Detection, Dimension Reduction, Data Storage and Data Query.
• Optimized Local Outlier Factor (LOF) algorithm by implementing K-D Tree to reduce time of searching K nearest neighbors. Reduce 89% run time on the test dataset.
• Used Redis Cluster as an LRU cache to reduce time of data query.
• Used Cloud Firestore as the NoSQL cloud database to store data for improving scalability.
• Built Docker Image for the web application to enable other developers to test and deploy easily.
• Developed interactive web pages using JavaScript, JSP and using Echarts to visualize users’ uploaded datasets.

May 2017 – Jun 2017 · 1 mo · Beijing City, China

• Developed a pipeline for processing and analyzing log data of the web server.
• Developed an ETL (extract, transform, load) process using RDD API and developed a Log Analysis Module of calculating Top-N popular modules of the website using SparkSQL and DataFrame API.
• Used JavaScript, JSP, HTML to develop interactive web pages and used Echarts to visualize the output of analysis module.
• Installed and configured Hadoop cluster.

Jun 2016 – Jul 2016 · 1 mo · Guiyang, Guizhou, China

• Collaborated with a team to develop a web application of Online Store using Spring, SpringMVC and MyBatis.
• Developed multiple modules, e.g. Product Management Module, Shopping Cart Module, Payment Module, and designed multiple reusable methods.
• Worked with APIs to connect the 3rd party platform Alipay, thus users can use Alipay to make secure purchases.