Model optimization - making models train faster on Nvidia GPUs

At Waymo, I developed a model surgery toolkit that automated tensor debugging and eliminated silent checkpoint restoration failures across large-scale training and evaluation pipelines, reducing debugging time by over 90% (from days to hours). I extended this toolkit to enable automated foundation model conversion between Waymo’s internal infrastructure and Google DeepMind’s Gemini training stack. Additionally, I implemented dataset checkpointing within Waymo’s foundation model pipelines using Grain, enabling fault-tolerant and resumable training. My work also included profiling tf.data and Python backends to guide throughput optimizations for large-scale distributed workloads.

At Tech Profuse, I developed an unstructured data extraction API powered by Gemini, processing over 50,000 bill of lading documents in 15 hours and reducing manual data entry by 98%. I also built a multimodal data extraction prototype by fine-tuning a LLaVA model with distributed training (FSDP/ZeRO) across 8 GPUs, significantly improving visual-text alignment accuracy. In parallel, I engineered a RAG-based support system leveraging Cohere LLMs for natural language issue querying, classification, and summarization—boosting support throughput by 130%.