I’m a Research Engineer at Google with a passion for building and improving cutting-edge generative models. My primary focus is on synthetic data, an area I’ve explored extensively through my work in both industry and academia. I believe synthetic data is a critical tool for scaling model capabilities and bridging performance gaps, particularly for specialized tasks.

From Theory to Impact: Scaling Gemini with Synthetic Data

At Google, I’ve had the opportunity to apply this expertise to some of the company’s most advanced large language models. From September 2023 to March 2025, I led the development of a data generation pipeline designed to create high-quality instruction-response pairs for fine-tuning large language models. This work was a direct contribution to Gemini, and the resulting data delivered an average performance boost of 0.03 across 25 languages on a standard multilingual evaluation set.

• I owned this project from end to end, creating a systematic experimentation flywheel to measure and validate progress.
• By executing over 50 fine-tuning experiments, I analyzed results to validate hypotheses and ensure performance gains were stable on next-generation models.
• I discovered and implemented a robust intervention that significantly improved data quality by leveraging more powerful models and advanced prompting techniques.
• To ensure the pipeline’s continued relevance for cutting-edge model development, I adapted it to work seamlessly with the latest models.

My approach to building these high-impact data pipelines is directly informed by my doctoral research, where I developed a first-principles understanding of how generated data can be used to probe and predict model behavior in novel situations. Foundations: Pioneering “Stress Tests” for Model Reliability

Foundations: Pioneering “Stress Tests” for Model Reliability

My PhD thesis, “Study of Data Augmentation for the Robustness of Deep Neural Networks,” investigated how models perform under dataset shift—when deployment data differs from training data. A key finding was that synthetic data could be used not just to make models more robust, but more importantly, to generate targeted “stress tests” at the evaluation stage.

The outcome was a new methodology for using cheap, generated data to systematically probe model behavior and accurately estimate performance drops on new, unlabeled domains. This insight has practical applications, such as guiding expensive labeling efforts toward areas of maximal impact. This foundational research equipped me with a unique perspective on using synthetic data to build more reliable and trustworthy systems—a perspective that continues to drive my work on frontier models today.