Oct 2023: Our work, "Towards Flexibility and Robustness of LSM Trees" has been accepted for publication by the VLDB Journal!
May 2022: Our work, "Endure: A Robust Tuning Paradigm for LSM Trees Under Workload Uncertainty" has been accepted for publication at VLDB 2022!
Aug 2021: Started new job as a Senior Applied Scientist at Etsy, Inc. on the recommendations ranking team.
Oct 2020: Our work "Learn to Earn: Enabling Coordination Within a Ride-Hailing Fleet" has been accepted at IEEE BigData, 2020.
Aug 2020: My work with Zillow Group on Fairness in Multistakeholder Recommendations has been accepted at FATREC workshop held alongside RecSys 2020.
Aug 2019: I will be visiting Anchorage, Alaska for KDD 2019. Hope to see you there!
Oct 2018: After spending a very worthwhile summer interning at Zillow, we have decided to continue our research together remotely.
May 2018: Excited to spend my summer at Zillow Group as an AI intern on the Personalization team.
Feb 2018: Our paper "Putting Data in the Driver's Seat: Optimizing Earnings for On-Demand Ride-Hailing" featured on The Morning Paper.
Jan 2018: Our paper "Markov Chain Monitoring" accepted at SIAM International Conference on Data Mining (SDM18), San Diego.
Dec 2018: Our paper "Impact of free app promotion on future sales" accepted at TSMO 2018: Workshop on Two-sided Marketplace Optimization
Oct 2017: Our paper "Putting Data in the Driver's Seat: Optimizing Earnings for On-Demand Ride-Hailing" accepted at WSDM 2018, Los Angeles.
This is an extended journal version of our previous work on developing Endure, a new paradigm for tuning LSM trees in the presence of workload uncertainty. In this work, we explore whether an expanded and more flexible design space called K-LSM can offer similar benefits as robust formulation of the throughput maximization problem that we used while developing Endure. We find that robust optimization is the only approach that consistently high performance of database systems in presence of uncertainty.
Traditionally, multistakeholder recommendations problems have been formulated as integer linear programs which compute recommendations in an offline fashion, by incorporating provider constraints. Such approaches can lead to unforeseen biases wherein certain users consistently receive low utility recommendations in order to meet the global coverage constraints. To remedy this situation, we propose a general formulation that incorporates provider coverage objectives alongside individual user objectives, in a real-time personalized recommender system.
Modern LSM-tree backed key-value stores co-tune merge policies, buffer sizes and the false positive rates for the Bloom filters across different levels of LSM-tree. These systems typically minimize the costs for fixed workloads. We augment them to make them robust to perturbations in workloads.
We combine the interpretability of vanilla reinforcement learning with combinatorial optimization techniques to propose a systematically tunable, scalable and effective framework to maximize earnings of a fleet of ride-share drivers.
Given an initial distribution of items over the nodes of a Markov chain, we wish to estimate the distribution of items at subsequent times. In deriving these estimates, we issue queries to retrieve partial information on the distribution of items.
We formalize the problem of devising a strategy to maximize expected earnings of ride-hailing service driver, describe a series of algorithms to solve the problem, and exemplify the methods on a large scale simulation of driving for Uber in NYC ... Read more.
Amazon's Free App of the Day program, aimed at improving app visibility using daily free promotions, is a compelling experiment in the 'economics of free'. We investigate its longer-term consequences on the performance of apps on Amazon Appstore.
My fullform CV is available here.