I build infrastructure that accelerates research velocity at Google.

As a Research Engineer, I specialize in designing systems that eliminate bottlenecks in LLM development. My work spans two complementary domains: infrastructure engineering and synthetic data research.

On the infrastructure side, I architect systems that decouple experimental logic from execution, achieving step-change improvements in research velocity. Currently, I’m building multimodal tool use infrastructure for Vision-Language Models (PaliGemma, Gemini, Gemma), enabling reliable visual instruction-following across computer vision tasks. Previously, I delivered 10x experiment acceleration (from ~10 to ~100 experiments per quarter) through consolidated, configuration-driven frameworks and scaled data loading capacity 15x by building robust TFDS infrastructure now serving 60+ datasets. I build systems where adding new experiments, datasets, or models incurs constant overhead rather than linear complexity.

On the research side, my expertise is synthetic data for post-training—specifically, how it scales to arbitrary domains when rigorous evaluation infrastructure enables tight iteration loops. As a core contributor to Gemini’s multilingual capabilities, I architected the end-to-end synthetic data pipeline that scaled instruction-following across 25 languages. The quality came from making experiments cheap enough to run 50+ fine-tuning iterations, letting systematic hypothesis testing surface the right interventions rather than relying on manual curation. I cover generation for Supervised Fine-Tuning (SFT) and preference learning, with deep knowledge of synthetic alignment methods (including RLHF/RLAIF limitations and alternatives) synthesized in my research series.

This dual expertise is grounded in my PhD research at Sorbonne University, where I developed methods using synthetic data to predict model behavior under distribution shift—principles I now apply to designing robust, production-scale systems that unblock researchers at frontier labs.

I write about infrastructure design and the shift from human to synthetic labeling at jmamath.github.io.