Amazon - Software Development Engineer
New York, NY
Generative Recommendations (GenRecs)
- Built foundations for a recursively improving recommendation-ranking model on a distributed service handling 7K TPS: data pipelines, tracing, and generative AI evaluation supporting a goal to cut irrelevance 50% by EOY
- Architected, load-tested, and delivered a Java/Guice recommendation-tracing framework across ~95% of processing paths; moved logs and embeddings off the hot path and streamed them via Kinesis Firehose and S3 with zero downtime
- Productionized a Java/Spark LLM evaluation workflow for 10K users per month via Bedrock, Step Functions, and EMR; integrated six scientist-defined prompts to flag an issue in a new experiment and verify a ~15% irrelevance drop
- Migrated two weekly Python/Spark data jobs end-to-end from a legacy platform to Scala/Spark on AWS EMR, Step Functions, and CDK; cut runtime from 48–72 hours to ~6 hours while processing tens of terabytes
- Automated on-call monitoring for TPS, size, and freshness across 90+ datasets in 23 marketplaces; surfaced recurring upstream failures, improving reliability and helping cut active operations tickets by ~15%
- Redesigned a React response-audit tool to visualize recommendation sourcing, filtering, and ranking per request across 30+ processors; enabled engineers and scientists to view responses, test changes, and debug on-call escalations
Shopping Guides (SG)
- Built personalization infrastructure and backend systems for Shopping Guides, powering product discovery across Amazon's homepage and 512 guide landing pages
- Owned the end-to-end Java/Spring design of a homepage embedding strategy; batched retrieval and concurrent dot-product scoring cut latency from >200 ms to ~160 ms; the experiment attributed $900K in sales and 3.91M clicks
- Built a concurrent Java/Spring backend for a multi-category homepage mosaic: retrieved, ranked, filtered, and assembled personalized results; the experiment attributed $1.2M in sales and 5.85M clicks over four weeks
- Collaborated in a cross-team referral attribution experiment for influencer affiliate links; replaced and standardized one-off logic with shared backend and UI tags, enabling attribution of $6.46M in sales and 12.31M clicks
- Implemented AWS Lambda category and product allowlists, blocklists, and quality checks, supporting the team's expansion from 304 to 453 guides; guide-page traffic rose 39.7% after launch
- Owned the upstream Java/Guice and CDK publisher for a cross-team, event-driven SQS workflow triggered after asynchronous Bedrock classification to support dynamic, LLM-generated Shopping Guides