Harsha Aduri

Experience

1. Applied Scientist Ii

   Amazon

   May 2017 — Present

   Leading effort on the building ML solutions to improve Seller experience in WW SPX Org (Selling Partner Experience Org) Currently working on Prompt-based learning for constrained hierarchical text classification for seller issues at different granularities. Worked on improving intent detection when the query is super short (less than 6 tokens). Introduced a novel dual-encoder transformer architecture and a combination of historical and current context from seller along with behavioral patterns prior to contacting us to improve accuracy. Worked on moving intent systems in PSAS to transformer based language models to vastly improve the understanding of the models as well as accuracy of the support systems. Worked on domain adaptation of large language models and further task adaptation to several downstream tasks like intent detection and task clustering. Worked on a granular intent recognition system (Selling Assistant) for long form Seller communications and predicting the best possible solution to Seller\'s issue thereby reducing the number of cases going to Associates for response and also the average handling time. Previously: Transparency (Product counterfeit detection, prevention and Brand protection) Buyer Risk Prevention (fka. Transaction Risk Management/Buyer Fraud Detection)