Building the intelligence behind real-world robots.

I build the systems that turn machine-learning research into robots that work in the real world. My focus is large-scale ML infrastructure and training optimization — the pipelines, data loaders, and schedulers that let foundation models like VLMs and VAMs train efficiently on massive multi-modal datasets and run reliably on-device.

Today I'm a Member of Technical Staff at Dyna Robotics, where I build the ML infrastructure and training-optimization stack that powers our foundation models — multi-modal Airflow pipelines processing thousands of video-hours per hour, cross-cluster distributed data loaders and intelligent workload schedulers that keep GPU utilization high across the training fleet, plus the on-robot inference path for low-latency trajectory execution.

Before Dyna, I spent five years at Amazon Astro in Lab126. A big part of that work was on the Floor Plan Algorithms team — building the computer-vision and geometry pipelines that generate floor plans from raw SLAM data, deployed to 8,000+ devices in production. I also architected a spatial query engine and DSL for programmatic floor-plan access, and later worked on robot memory, summarization, and LLM-driven action planning for next-generation Astro.

I have an MSE in Computer & Information Science from Penn and a B.Tech in IT from VJTI Mumbai. My published research spans word-embedding debiasing, semantic relatedness, computer vision, and reprogrammable signal-processing hardware.