* Query Slot Detection for Microsoft Enterprise Copilot (M365 Chat)Led the development of an ML model for intent and slot detection in Microsoft Enterprise Copilot queries. Managed the entire development cycle, from ideation and data curation to model training and deployment. The model has shown significant potential in improving the relevance metrics for the query entity suggestion service in Copilot.*Learned Ranker Models for Enterprise Entity SuggestionsDeveloped and deployed ranking models for entity suggestions (Emails, Files, Keywords, etc.) in query formulation for Microsoft enterprise solutions like Outlook and Teams. These models significantly enhanced customer engagement metrics, such as time-to-success, driving increased revenue for the company.* Document Attachment Suggestions in OutlookDeveloped a semantic similarity model to improve document recommendations in Outlook's Suggested Attachments feature. Leveraged Turing-based architecture to estimate contextual relevance, integrating the similarity index into the ranking model. This approach led to significant improvements in ranking metrics.

Ad Relevance Estimation Using Distributed and Local RepresentationsBuilt a deep learning model to predict the relevance of product advertisements based on searchqueries. The model combined semantic match signals between the query and ad with exact tokenmatches and incorporated side-information to enhance relevance scoring. Utilized a duet archi-tecture with CNNs to fuse semantic and syntactic matches, implemented using TensorFlow, anddeployed on AWS SageMaker.

Ad Click-Through Rate Estimation ModelDeveloped a click-through rate prediction model for product advertisements. Proposed a systemwith separate models for the top and bottom ad widgets on the page, significantly improvingrevenue. The initial rule-ensemble model was later upgraded to a deep learning-based approach.Conducted a comprehensive feature importance study to optimize model performance, storage,and speed.

Speech processing and applied Machine Learning.