Ads Ranking

0 -> 1 ML for the short video platform Josh- Bootstrapped and maintained model training infrastructure- Transformer models: Retrieval (two-tower); Ranking (Multitask); LLM-like models for long content history based user models (Pinnerformer-like); Neural blending & diversification, etc- Vision-language models (BLIP2-like)

Machine Learning- Lead for collaboration between Uber AI and Uber EATS organizations: E2E Recommendation system for UberEATS including Transformer architectures for ranking, Two Tower Embeddings (TTE) training framework, Graph Candidate Generation algorithms like PPR, etc - Led projects at the intersection of ML Infra and ML algorithms for UberAI (Michelangelo)Algorithms- Novel vector search algorithm in Lucene using LSH (SimHash) and the TMOR operator (WAND query)- Performant geospatial embedding JAVA library using s2 cells and random projections (ETAs)- Novel map matching architecture using Kalman filters that vastly improved the state of the art at Uber (JAVA Backend) - Part of team responsible for developing and maintaining a JAVA sensor fusion library for location improvement on Android

Modeling lead for maintaining a vertical of language understanding models for Alexa (from incubation to launch) and also worked on search and entity resolution

- Performant word and graph embeddings using SVDs (published in NeurIPS '15) -- FastEmbed- Inference from time on Twitter: Proposed and evaluated a system for understanding the temporal evolution of topics on Twitter and for quantifying a user's interest in a topic from his/her tweet times. Investigated the value of user overlap for understanding the flow of topics with time on Twitter feeds.- Evaluated channel estimation and tracking algorithms for beamforming and tracking with compressive measurements / random projections. Gave a simple characterization of the design space made available by random projections using estimation theoretic / statistical bounds- Teaching assistant

Proposed algorithms and conducted simulations

Implemented computer vision algorithms (medical imagery)