Han Xia

About Han Xia

TL;DR: Software Engineer specialized in architecting large-scale Deep Learning Platforms (1B+ daily predictions,$500M+ business impact) at Amazon Last Mile.(Resume is attached under Featured section)I am a Software Engineer with a proven track record of architecting and scaling production-grade Deep Learning Infrastructure and MLOps ecosystems. During my tenure at Amazon, I served as a Founding Engineer for the Last Mile Time Estimation service, where I led the zero-to-one architecture of distributed systems processing 1B+ candidate routes daily to optimize delivery promises for ~52M packages.My expertise sits at the intersection of Distributed Systems, High-Concurrency Backend, and Machine Learning Infrastructure. I specialize in industrializing deep learning at scale—transforming complex research models (LSTMs) into resilient, high-throughput platforms that operate under strict real-time SLAs. I have a consistent history of driving massive business impact, generating over $500M in annualized operational savings through systemic performance breakthroughs and architectural innovations.

Experience

1. Software Engineer Ii - Last Mile Time Estimation

   Amazon

   Jun 2022 — Present · Bellevue, WA, US

   Deep Learning Delivery Time Estimation Service | Founding Engineer • Architected a zero-to-one distributed deep-learning-driven inference engine (AWS ECS, SageMaker) deeply integrated into Amazon Last Mile route planning system, processing 1B+ daily route evaluations for ~52M packages while sustaining a burst throughput of 10M+ requests in 2 minutes.• Generated $500M+ in annualized savings by improving Planned vs Actuals delivery time accuracy (MAPE improved by ~100bps, IQR reduced by ~23 mins) across 600+ stations through the large-scale deployment of PyTorch-based LSTM models.• Enabled the feasibility of heavy LSTM models in route planning system by minimizing inference load by ~95% via a \"decompose-and-stitch\" strategy, aggregating reusable zone predictions to derive estimates for millions of combinatorial route permutations without redundancy.• Engineered a \"Weather-Aware\" routing inference capability that drove a ~14% reduction in heat-related incident rate (0.723 to 0.619 YoY) even with a +15M hour increase in on-road exposure and record-high heatwaves.Please check Projects section for two other projects Real-Time Inference Latency Optimization | Lead Engineer Enterprise MLOps Platform & CI/CD/CT Strategy | Lead MLOps Engineer

Education

• Zhejiang University

  Bachelor of Science, Biology, General

• Brandeis University

  Master of Science, Computer Science