Leading the Autopilot Maps team, where we are creating high-accuracy maps for an ever growing fleet of Tesla vehicles to enable full autonomy with geospatial data from within and beyond the cars' immediate surroundings. Our fleet learning methods help Autopilot have an increasingly detailed knowledge on how to navigate the road ahead.

* Worked on time-of-arrival models to estimate driver travel time in cities, resulting in siginificant increases in prediction accuracies accross the board and geographies. The improvements help make driver dispatch and fare pricing more efficient, and result in better passenger experience.* Developed a granular passenger demand prediction framework to anticipate changes in the volume of passenger requests in the near future. Prediction is essential to mitigate the effects of on-demand ("prime time" or "surge") pricing by allowing us to better distribute available supply over all areas ahead of time.* Developed a driver prepositioning system in conjuction with demand forecasting to implement long-range driver dispatches, optimizing for passenger request fulfillment. Rebalancing cars from remote areas as a response to expected future rider requests further decreases supply-demand misalignment, increases utilization, and creates a more efficient marketplace.

Technical lead for a team of data scientists studying online social networks, statistical user behavior, and Twitter content. Focuses on describing and predicting user engagement, modeling graph connectivity, and on building frameworks that enable this research. The purpose of these studies is to understand long-term Twitter usage trends and to help make strategic business decisions.Author of PyCascading, an open-source Hadoop framework for running MapReduce jobs purely in Python using Cascading: https://github.com/twitter/pycascading

* Predicting the popularity of online content. Using hit data from Twitter, Youtube, and Digg, developed a model to predict aggregate user behavior that gives rise to different download rates of online content. Showed that initial access rates of content can be used to forecast long-term user interest.* Friendlee: a mobile application for intimate social ties. Developed an Android application that replaces a mobile phone's traditional phonebook by analyzing the users' calling and location patterns building an intuitive, weighted social network. Built the server backend as well.* Relationship types and diversity in online communities. Analyzed the role of link semantics and long-tail user trends in an online community website that emphasizes discussions and group formation around current topics.

Network analysis of human protein interaction networks. Studied the role weak and strong links play in information search in mobile social networks.

Worked with large social networks data from mobile phone communications in Albert-Laszlo Barabasi's lab. Developed an accurate churn prediction algorithm and modelled mobile service adoption and diffusion among the users.