At Scale AI, I designed PyTorch experiments for supervised fine-tuning, reward modeling, PPO, and DPO, and built distributed post-training pipelines around expert preference rankings. My work accelerated RLHF iteration cycles 40%, lowered compute overhead 22%, and improved target-model win rate 18% versus baseline checkpoints.
I also architected a petabyte-scale multimodal data curation engine and led data-preparation workflows for video, audio, image, and text. At Databricks, I built retrieval and forecasting systems, improving SKU-level forecast accuracy 18% and reducing the deployment lifecycle from six weeks to 45 minutes.
