Applied scientist in Amazon Personalization.

Windows Copilot, new features in Bing Chat, Text Prediction in Edge Browser and inference optimization of large language models. Working on Large Language Models, their applications and inference optimization, at Microsoft Turing. https://turing.microsoft.com

Sequential decision making from noisy and correlated observations in the context of recommendation systems, A/B testing, experiment design etc. Thesis: Structured and Correlated Multi-Armed Bandits: Algorithms, Theory and Applications.- Developed a novel framework to sequentially select the best action from a set of available actions, where the rewards corresponding to different actions are correlated and noisy. - Proposed novel online learning algorithms that exploit the knowledge of correlations. - Analyzed the algorithms theoretically and empirically through experiments on recommendation system datasets.- Applied this work to study the problem of recommendation systems, resource allocation, scheduling systems and the problem of client selection in Federated Learning

Selecting best available ML model for a user's query at Amazon Search while accounting for peculiar customer behavior on Amazon (modeling the fact that users may purchase items after conducting multiple refined searches.)Project: Empirical MDP based modeling to capture session-aware customer-amazon interaction Paper: Bayesian Regularization of Empirical MDPs. - Proposed a novel methodology to decide the best available search algorithm for the search query typed in by the customer- Modeled the customer's interaction with the Amazon's search algorithm as a Markov Decision Process using historical data to account for the fact that a user may conduct/refine their search upon viewing the set of results.- Proposed novel regularization based solutions to the empirical MDP and evaluated their performance on pre-existing search log data- Our proposed solution effectively combines the existing search algorithms by using different algorithms for different search queries and show significant performance gains over each of the individual algorithm

Uncertainty aware metrics and safety oriented predictions for autonomous vehicles at Uber ATG.Project: Uncertainty Aware Failsafe Predictions for Traffic Actors around Autonomous Vehicles- Worked on evaluating the performance of mainline prediction, that predicts the trajectory ofactors around the self driving vehicle. - Incorporated new performance metrics that account for the uncertainties present in the prediction- Designed a safety oriented deep learning model for trajectory prediction, that activates when the mainline prediction’s performance is below par.

Estimating bond curve parameters using numerical optimization at Morgan Stanley Strats and Modeling.Project: Trust Region Optimization for Estimating Bond Curve Parameters• Designed a minimizer for estimating bond curve parameters using numerical optimization• Reviewed literature for Line Search and Trust Region Optimization techniques• Modified and Implemented DogLeg trust region method to develop a minimizer• Introduced Broyden’s method to update Jacobian for reducing function evaluation counts• Produced results with same accuracy in 90% less evaluation counts compared to NAG minimizer,Implemented Warm Calibration to further achieve 98.6% less evaluation countsIt is being used in production as a generic minimizer library for multiple different applications

Worked with Prof. Emanuel Popovici in the Embedded systems group at UCC. Project: Synthesis of Reliable Circuits from Unreliable Components• Constructed an AND - Invert Tree model to represent digital combinational circuits consisting of unreliable inputs and components• Devised an algorithm to synthesize AND-Invert trees using Boolean Matching and AND Masking techniques, Optimized algorithm to produce circuit within given reliability constraints• Analyzed the errors in input that the designed circuit can tolerate by estimating the reliability• Developed an algorithm to evaluate output error probability for AND-Invert model of circuits