At Databricks, I trained and evaluated large transformer models on multi-node GPU clusters, owning the pipeline from data interfaces through checkpointing and evaluation. I implemented hybrid sharding and mixed-precision training for a 13B-parameter model, improving throughput by roughly 35% without adding nodes.
At H2O.ai, I fine-tuned open-source LLMs for text classification and summarization. I also moved a GNN workload to multi-GPU training, cutting epoch time from several hours to under one, and introduced a framework for more reliable model comparisons.
Earlier, at Domino Data Lab and DataRobot, I built and maintained training, evaluation, and serving workflows. My work included improving checkpoint recovery, optimizing inference prototypes, and fixing data and reproducibility issues in production pipelines.
