Michael Lee

Staff Software Engineer with 10+ years building large-scale recommendation systems and ML infrastructure.

Yahoo

News owns the ranking and personalization stack serving tens of millions of monthly active users, from candidate retrieval through real-time inference. Proven track record of technical leadership, cross-functional platform strategy, and shipping systems that measurably move engagement and retention.

Staff Software Engineer - Content Ranking & Personalization - Yahoo News at Yahoo (2024-01 – Present)

• Led the design and evolution of the multi-stage recommendation pipeline, built in Python and Java/Scala on Apache Spark, powering personalized feeds across Yahoo News web, mobile, and native app surfaces for tens of millions of monthly active users.
• Architected a real-time feature store using Kafka, Flink, and Redis that reduced feature freshness latency from minutes to under 5 seconds, enabling same-session personalization based on in-session reading behavior.
• Redesigned the candidate retrieval layer using approximate nearest-neighbor search (FAISS) over article embeddings generated with PyTorch, increasing recall of relevant long-tail content by 34% while holding p99 latency via a gRPC-based serving layer.
• Drove migration of the ranking model from XGBoost/LightGBM gradient-boosted trees to a two-tower deep learning-to-rank architecture in TensorFlow, improving click-through rate by 11% and session depth by 7% in production A/B tests.
• Led migration of the ranking and serving infrastructure from GCP to a Kubernetes-based platform on AWS (EKS) using Terraform and Helm, cutting new model deployment time from days to hours and reducing compute costs by ~20% through autoscaling and right-sizing.
• Built an internal React/TypeScript observability dashboard (using Recharts and D3 for visualization, backed by a Node.js/GraphQL API) for ranking experiments and feature drift detection, reducing time-to-insight for other teams by ~40% during A/B test analysis.
• Set technical direction for the ranking platform's roadmap, partnering with Product, Editorial, and Data Science leads to balanced engagement, diversity, and editorial-quality objectives in the ranking function.
• Built the team's offline evaluation and experimentation framework in Python (pandas, scikit-learn), orchestrated with Airflow, cutting the average time to validate a new ranking signal from two weeks to three days.
• Serve as technical mentor for a pod of 6 engineers; run design reviews and set coding and architecture standards for the recommendations codebase (Python, Scala, Java).
• On-call owner for the ranking and serving infrastructure (Kubernetes, Prometheus/Grafana for observability); led incident response and root-cause postmortems for personalization-related reliability issues, improving service SLA from 99.9% to 99.97%.

Senior Software Engineer, ML Platform - Yahoo News at Yahoo (2021-01 – 2024-01)

• Built the online model-serving layer (TensorFlow Serving on GCP/GKE) used to deploy ranking and personalization models to production, supporting shadow deployment, canary rollout, and automated rollback via Istio traffic splitting.
• Developed the user embedding pipeline (Spark + Python, implicit feedback from clicks, dwell time, and shares) that became a foundational feature set adopted by three downstream recommendation surfaces beyond News.
• Partnered with Data Science to productionize experimental ranking models (scikit-learn/TensorFlow → production Java/Scala services), reducing the research-to-production cycle from roughly 6 weeks to under 2.
• Introduced a real-time A/B experimentation framework for ranking changes, built on Kafka event streams and a custom stats engine in Python (SciPy), giving the team statistically rigorous, faster read-outs on engagement impact.
• Contributed frontend components (React, Redux) to the internal experimentation console used to configure and monitor live ranking A/B tests.

Software Engineer, Search & Personalization at Yahoo (2018-01 – 2021-01)

• Built backend services (Java, Spring) for personalized search result re-ranking, integrating user history signals into the core ranking pipeline.
• Designed and implemented a distributed logging and feature-extraction pipeline (Kafka, Hadoop/Hive) processing billions of daily events for downstream ML training.
• Contributed to the migration of legacy batch-scored recommendations (Hadoop MapReduce) to a near-real-time scoring architecture (Spark Streaming).
• Built lightweight internal UI tools (JavaScript, React) for engineers to inspect and debug personalized search rankings.

Full Stack Developer at Verizon Communications Inc (2017-01 – 2018-01)

• Integrated frontend UI/UX enhancements with backend services, resulting in a 30% increase in user engagement metrics for My Verizon and enterprise dashboards.
• Optimized database queries and implemented caching strategies (Redis/Memcached) that decreased page load times by 40% and supported peak traffic during high-volume periods.

Bachelor of Science in Computer Science – University of Maryland Baltimore County (2012 – 2016)