Brandeis University

University of Rochester

Vassar College

Computer scientist at large

I founded the Machine Learning Research team at Medium and supervised and lead a team of ML engineers and scientists solving challenging problems in recommender systems and natural language processing.

Medium

Chief Scientist

Principal Scientist

Customized probabilistic tensor factorization to work on Pandora's rich dataset of 85 billion thumbs and implemented the custom algorithm in Spark and Cython. Used the implementations to generate high quality embeddings (i.e. vector representations) for most objects in Pandora's musical universe, including artists, albums, tracks, playlists, and stations. The embeddings, in effect, allow Pandora to recommend anything from anything, where "anything" includes items in the vast long tail. Lead, consulted, or worked on several crucial projects where the embeddings or derivatives thereof played a central role, including:

• Personalized song recommendation; most impactful of over 80 recommenders (Team lead)
• Automatic hierarchical genre tagging of artists, albums, tracks, and stations (Team lead)
• Diversified recommendation of expert curated playlists on the Browse page (Team lead)
• Automatic playlist extension
• Long tail song recommendation in Pandora Radio
• Contextual bandit based row ranking in the Browse screen (consulted)
• Cold-start song testing via deep learning on spectrograms aka spec2vec (consulted)
• Automatic generation of mood and genre based playlists  (consulted)
• Artist direct messaging (consulted)
• Automatic generation of group playlists for Premium Family accounts (consulted)

Pandora

Staff Scientist

Derived, prototyped, and implemented large scale machine learning systems that personalize song recommendations using Pandora's unique dataset of over 85 billion thumbs. 

• Pandora's core radio: Derived and implemented the most impactful of over 70 recommenders in use at Pandora.

• Station and Song Embeddings: Developed a probabilistic ML pipeline that processes Pandora's dataset of 85 billion thumbs to generate high quality embeddings for stations and songs. The embeddings are used  in multiple data products across Pandora

• Autoplay and Playlist Extension : Prototyped the autoplay (play similar songs after a playlist has completed) and playlist extension (add similar Songs to a playlist) experience for Pandora Premium (Pandora's on-demand offering)

• Thumbprint Radio: Deeply involved in the design and implementation of Pandora's most popular music station

ML techniques: Large scale matrix and tensor factorization, clustering, regression, probabilistic methods, transfer learning 

Technologies:  Spark, Hive, PyData stack, Angular (for front-end product prototyping)

Languages:  Python, Cython, SQL, Scala,  Java, Javascript, Typescript

Senior Scientist

Zite was a personalized reader for mobile devices that learned your interests and delivered the best of your favorite magazines, newspapers, authors, blogs, and videos.

• Collaborative Filtering Improvements: Improved the quality of Zite's recommendations by conceiving of, implementing, and cross-validating changes to Zite's large-scale matrix-factorization based recommender system (indexing over one million new documents per day).

• Url Canonicalization: Identical webpages with different urls confound collaborative filtering algorithms and waste resources. Solved this problem by designing and implementing a system that learns rules for canonicalizing urls of a given http domain. Deployment of rules resulted in a 32% decrease in documents stored.

• A/B Testing Framework: Implemented a framework that allows Zite’s engineers and project managers to measure the effect of changes to backend machine learning algorithms on user engagement (measured by the number of documents viewed, session length, etc.) and user satisfaction (likes/dislike ratio). Split testing results are viewable as graphs in mixpanel (mixpanel.com) and are segmentable by all previously defined segmentations.

• Content Extraction Regression Testing Framework: Designed and developed a framework to regression test changes to Zite’s content extraction algorithms across the 10,000 most popular domains. Used the framework to safely resolve numerous content extraction issues. 

Methods / Languages / Technologies : Model based Collaborative Filtering, Named Entity Recognition, Classification, Regression, Boosting, Learning to rank, Python, Cython, C/C++, Eigen, OpenMP, Numpy, Scipy, Pandas, Scikit-Learn, MySQL, Cassandra, Linux, EC2, S3

Zite

Senior Research Engineer

• Designed and developed web based tools that that allowed product managers to test different pricing models for virtual goods in Zynga’s games. Ran A/B tests to verify and quantify revenue increases in international markets. 

• Automated the tuning of pricing models to further optimize revenue. 

Technologies used: Java, PHP, Spring, Maven, Hibernate, MySQL, MongoDB

Zynga

Software Engineer

Involved in the design, development and day to day operation of the Catalog Manager component of Amazon Webstore (webstore.amazon.com). Implemented architectural changes that promote greater scalability, flexibility, and stability of this central, highly-available component.

Keki Burjorjee