In this talk we are going to provide details on a new ray-based solution for a large-scale recommender system for Josh, the leading short video platform in India having 85M DAU, 170M MAU and 100M+ Videos. Our design is motivated by the progress in large language modeling and comprises a single big model with sparse inputs and outputs with both early and late fusion for different outputs (like retrieval and ranker outputs). These models are scaled with a universal graph search algorithm which is a generalization of HNSW. We will discuss the advantages of this approach, such as simplified vertical integration and end-to-end content training from raw user actions and present empirical results. This solution led to double digit improvements in engagement metrics on the platform
Dinesh Ramasamy is an ML engineer at Josh based in San Mateo. He was formerly a Senior Staff Software Engineer at Uber AI. Dinesh has experience in building ML Infrastructure and Modeling solutions for Recommender Systems at both Josh and Uber. Before ML, Dinesh worked on improving Location accuracy at Uber AI and Uber Maps. Dinesh holds a Masters and PhD in Signal Processing and Wireless Communication from the University of California Santa Barbara and a Bachelors in Electrical Engineering from IIT Madras.
Come connect with the global community of thinkers and disruptors who are building and deploying the next generation of AI and ML applications.