Mountain View, California, United States
As a Staff Software Engineer at Google Research on Gemini Pre-Training, I build the data infrastructure and quality-control pipelines behind Google's next-generation Gemini and Gemma models. My work sits at the foundation of these models: near-duplicate detection, poisoned and low-quality data filtering, and large-scale dataset curation that directly improves model quality and compute efficiency. As a Gemini Data Captain, I've led pre-training scaling runs backed by significant TPU compute and contributed to critical public releases. Before this, I spent nearly five years in Search Ads Quality, where I built and scaled LLM-powered products from beta to general availability that reached 20% of Google's advertisers and drove billions of dollars a year in incremental revenue — work called out in Google's earnings reports and Google Marketing Live from 2023 to 2025. I worked across the full LLM lifecycle there, from latency-critical real-time inference serving a billion queries a day, to post-training with SFT, RLHF, RLAIF, and chain-of-thought prompting. As tech lead, I set technical direction, ran red-teaming, and mentored engineers. My path to LLMs ran through large-scale data and ML: building graph-AI product-clustering infrastructure for Google Shopping that processed petabytes to graph billions of products, and earlier, data and ML pipelines at Media.net. I got my start as a competitive programmer (ACM-ICPC Asia regionals) and still bring that algorithmic instinct and strong distributed-systems fundamentals to everything I build. I like the hard, high-leverage problems that make large systems actually work. Always happy to connect with people building LLMs, data infrastructure, or large-scale ML.
- Gemini Pretraining Data Quality @ Google Research/GDM
• Built and scaled several Search Ads AI products (AI Max, Automatically Created Assets) from beta to GA, reaching 20% of Google’s advertisers and driving $XX B/yr in incremental revenue, called out in Google Earnings(2023-25), GML (2023-25) • Led LLM post training for text search ads with objectives like long term revenue & personalization using techniques like SFT, RLHF, RLAIF & COT Prompting, built the first Google product to drive incremental revenue from generative AI without HIL. • As tech lead, set technical direction and roadmaps, ran red-teaming, mentored SWEs and interviewed candidates.
* Solved frontier AI challenges to build latency sensitive LLMs to generate factual, revenue driving, query relevant ads in realtime (SLO 1ms) for billion search queries/day. * Built Campaign Assets for Responsive Search Ads from Beta to 1M+ Customers.
Data Team, Apps Division • Daily responsibilities include designing infrastructure and improving big data pipelines. • Reduced processing time of hourly ETL processes by 40%, reduced resource requirements and data storage space. • Built a machine learning pipeline to predict life time value and churn rate of a user to help the business analytic. • Developed a monitoring tool to analyze website’s availability, performance, user experience and send alerts.
Part Time - Contributing algorithms and data structure related problems which helps students and programming enthusiast learn new tips and tricks in programming world.
Designed and developed a Banner Management System to run intelligent ad campaigns and serve banner ads through various Click-through rate optimising algorithms and machine learning models. • Implemented and tested multi-armed bandit algorithms package that improved click through rate by 32%. • All code was reviewed, perfected and pushed to production. Received a full time job offer.
• Curated 50+ problems for the company's new online judge. • Curated and Managed three major monthly contests. • Conducted teaching sessions to explain the approaches to solve contest problems.