Feb 2018 – Apr 2022 · 4 yrs 2 mos · Mountain View

• Staff Research Engineer and Tech Lead/Manager in the DeepMind Applied team, deploying ML research and techniques to solve Google product problems. Working (and have worked) in building ML infrastructure and applying ML to a variety of domains, e.g., sales efficiency, scheduling optimization, speech synthesis, etc.

Aug 2012 – Feb 2018 · 5 yrs 6 mos · Mountain View

• Senior engineer on a large-scale machine learning platform, applying semi-supervised learning and deep learning to various products (conversation understanding, image understanding, search). Exploring ways of training deep networks with sparse training data, and compressing models for on-device computation.
• Previously built large-scale test infrastructure to run integration tests of Google's Display Ads serving stack.

Sep 2011 – Aug 2012 · 11 mos

Jun 2011 – Aug 2011 · 2 mos · Greater New York City Area

• Programmatic Log Verification:
• Designed and implemented a log verification module that lets test writers programmatically check test outputs
• Augmented the existing log-diffing module to provide a cleaner HTML output for comparing generated logs with golden logs
• The new verification and diffing modules reduce the effort necessary to debug output

Oct 2010 – Jun 2011 · 8 mos · La Jolla

Jul 2008 – Jul 2010 · 2 yrs

• Promotions Optimization (Frozen Food Retail Firm)
• Analyzed and determined efficiency of product promotions for a frozen food retailer through linear regressions
• Developed a mixed integer programming engine in C and Python which creates an optimal schedule of product promotions while satisfying constraints of budget, time, product availability etc.
• Deployed a Tomcat webapp on an Apache server as a frontend, with a SOAP service connecting the webapp to the Python/C linear programming backend
• Optimized promotion schedule is expected to deliver savings of 15% by improving overall ROI on trade spend
• Recommendation Engine (Retail Firm)
• Built a recommendation engine in C, MySQL and Python that uses nearest-neighbor and neural network models to make product suggestions to 3 million customers each week
• Automated the system to run without any human intervention, coupled with a monitoring system to send alerts, current status mails and system logs; automation was essential as the engine was signed for a 3-year contract
• Deployed BIRT and Flex-based reports that generated performance reports for salesmen, depots and regions
• Weekly product suggestions have generated an approximate 5% ($70 Mn) increase in revenue in the first year
• Netflix Challenge
• Developed a proof-of-concept SOAP web service powered by a combination of SVD and neural networks that recommend movie titles for a user based on his viewing history
• The system analyzes in real-time a database of over 25,000 movies and suggests top `N' movies that satisfy user-specified filters (genre, cast, censor rating)

May 2007 – Jul 2007 · 2 mos

• Object Tracking – R&D
• Developed a video player aimed at “video advertising”, which would allow a user to select and hyperlink objects in videos to product URLs
• Selected object was tracked in the video and the hyperlink mapped to the area covered by the object in each frame
• Object tracking was implemented using a variation of the Conditional Density Propagation algorithm
• Increased robustness of the tracking system by accounting for duplicate objects, partial occlusions, changes in shape