At Reddit, I built the production multimodal Media Annotation pipeline, turning 400K+ images and 30K videos per day into structured semantic features for recommendations, search, and moderation. I also built human-reviewed evaluation datasets and compared multimodal models for quality, throughput, and cost.
At Reddit, I engineered retrieval and ranking systems, including multi-task two-tower retrieval that improved recall@2000 by 13% over the baseline. I also migrated pCTR and pCVR models to deep neural networks, achieving a 2–4% CTR lift while beating the previous latency baseline.
At Google, I contributed to Apache Beam and Dataflow distributed processing, including cross-language Python pipeline execution and BigQueryIO streaming writes accelerated up to 3x. Earlier, I built streaming feature systems at Reddit and transformation operators for Workday Prism Analytics.
