== Search Team ==

August 2017 to April 2019

* Created MVP then fully productionized system to predict business properties from keyword mentions in reviews.

* Used learning-to-rank to train regression models for ranking businesses, released and A/B tested the models online.

* Migrated unspecialized searches from deprecated homegrown system to elasticsearch.

* Improved team tools: ranking model training pipeline (better data persistence, optional use of Spark); qualitative recall/ranking comparison tool.

== Spam and Abuse Detection Team ==

August 2013 to August 2017

* Developed and maintained Yelp's centralized service that synchronously detects disruptive or abusive content submitted by users, using a DSL of rules.

** Major contributor to initial implementation, deployment, migration of existing functionality, validation, documentation, logging, and monitoring.

** Advised on subsequent major changes to the service, including an asynchronous mode and experimentation support.

* Developed and productionized specialized review quality classifier for businesses without much content.

* Improved infrastructure for spammy or low quality review detection.

** Introduced Redshift to store data too large for MySQL, cluster soon became a data warehouse and new computation engine for the team.

** Led a project using Kafka to asynchronously move values of signals as they are computed into multiple datastores (Redshift, Cassandra).

** Reduced redundant computation of signals when no new data is available.

Developed a system for word sense disambiguation in medical corpora (specifically clinical notes). The algorithms I used and compared varied between less successful graph-based approaches and more successful topic modeling.

Interned on the risk engineering team detecting payment fraud.

My work revolved around the boundary of the team's payment fraud detection service and its clients. One project created a minimal in-memory implementation for use as a fallback by clients upon service failure (e.g. timeout). Another sent the service's responses to an internal publish/subscribe system, with multiple subscribers from other risk engineering team applications.