Aravind Vasudevan

Background

About Aravind Vasudevan

I am an engineering leader specializing in the reliability of large-scale Generative AI systems. I lead the teams that ensure cutting-edge AI products like Google's Gemini are stable, scalable, and ready for millions of users worldwide. With over 15 years of experience at Google, AWS, and Synopsys, I solve complex challenges in AI infrastructure, distributed systems, and production engineering. My passion is applying foundational principles to architect next-generation platforms that power high-impact products. As a leader on Google's Gemini and AIStudio SRE team, I am accountable for the products' end-to-end health, setting the multi-year technical vision for AI Reliability. I drive cross-functional initiatives, secure senior leadership sponsorship, and define foundational reliability patterns for new AI systems. I am currently architecting the deployment strategy for Google's agentic AI platform, leading its productionization and adoption across multiple Product Areas. My work also focuses on accelerating AI product velocity through clever client-server protocol opimizations. I've pioneered open-box monitoring solutions for LLMs to eliminate manual outage detection, addressing ~90% of the root causes from prior production incidents. I led a cross-functional effort that reduced end-user latency by ~220ms in key markets through targeted network optimizations. Previously at AWS Lambda, I helped develop and maintain a petabyte-scale data warehouse processing data from the Lambda service fleet. My contributions included:(1) improving data completeness to 99.9%+ with sophisticated retry mechanisms,(2) reducing the cost of Spark jobs on EMR fleets through ML-based capacity forecasting, and (3) developing self-service capabilities for customers to generate multi-dimensional reports and dashboards. At Synopsys, I applied my research in algorithms and machine learning to the Electronic Design Automation (EDA) domain. I introduced a method of trading accuracy for execution time by integrating a machine learning model into a core graph search algorithm. This achieved a ~10x improvement in runtime with less than 2% loss in accuracy and was immediately adopted by customers into production workflows. I also made significant performance improvements to the distributed task execution environment for static timing analysis. My approach is built on a strong research foundation, including a PhD in Computer Science and postdoctoral work in deep learning model optimization, and I thrive on leading teams to solve the industry's most critical reliability challenges.