Troy Peng Email & Phone Number

California, United States

Seattle, Wa, Us

- Contributed to the EU expansion of Rufus, the Amazon Store GenAI chatbot, enabling its deployment to transform the online shopping experience for European customers.- Rearchitected the Amazon Search QU Deep Learning model service pipeline, enabling the team to scale out model deployments and empowering partners to manage their own pipelines. Ultimately improved search capabilities and elevated the user experience across global markets.- Key contributor in building a cutting-edge ML training library from scratch, built on top of PyTorch Lightning and HuggingFace, and fully integrated with MLflow, providing a scalable and efficient model training experience.- Released key features of the ML training library used by all Applied Scientists across team and partners, enabling all users to run inference & soft-label generation on AWS Batch and SageMaker.

Seattle, Wa, Us

- Set up the production model refresh pipeline for a 100 Million Parameters LLM Model in the Canada marketplace, delivering a 8-figure incremental annualized gain in GCCP (Growth-Adjusted Composite Contribution Profit).- Developed and maintained a production-level model inference pipeline to improve DP-CTR (Amazon Detail Page Click-Through-Rate), ensuring efficient processing of 2000M throughput per day.- Conducted in-depth research to utilize ONNX and TensorRT to significantly optimize and accelerate model inference, improving overall system performance and reducing latency in high-throughput environments.

Global, Cn

• Architected and developed feeds and friends’ recommendation back-end systems for DiDi 99 Payrecommendation engine, with Java-SpringBoot and Golang-Gin.• Established debug APIs and logging tools for friends’ recommendation back-end service, and utilizedRedis as cache system which reduced average recommendation API latency from 30ms to 20 ms• Debugged and solved SQL latency overhead, reducing the avg SQL I/O latency from 12 ms to 3.68ms, 98 Percentile latency from 30ms to 10ms• Refactored recommended feed filter to benefit from Redis concurrency to reduce latency overhead when filtering large number of feeds, and implemented debug APIs to manage Redis via http request• Implemented new flow engine of the feed recommendation back-end system for decoupling and higherscalability, and enabled cache system for personal candidate feeds and global candidate feeds.• Used JIRA and Confluence to keep track of project development progress, bugs and issues, and utilized GitLab for version control.• Developed and maintained unit tests for feeds and friends’ recommendation back-end services, and used CICD tool for continuous integration, SIM environment testing, and deployment to Production environment.• Took over Carpool-Pricing backend and performed routine inspections and system maintenance.

Champaign, Il, Us

Research in development of Centralized ACLS Cardiopulmonary Resuscitation Guidance System.

Haidian, Beijing, Cn

Testing Autonomous Warehouse Robots navigation algorithm with SLAM, Radar, and Camera.

Master'S Degree, Computer Science

Bachelor'S Degree With Honors, Computer Science

High School Diploma