Backend product development bridging smart glasses with Instagram

Developed MLOps infrastructure consisting of Kubeflow pipelines on Azure AKS, ACR, and SQL Server to deliver periodic prescriptive analytics for higher ed clients.

Designed reusable pipelines for feature and target creation, model building, and model scoring with model-specific configurations using Python, Pandas, Scikit-learn, and Docker, which scores ~1GB of real-time data every 30 min.

Cleaned and processed model data by identifying target leakage and created markov models to analyze data availability trends in dynamic raw data, resulting in more representative modeling datasets.

Introduced new testing infrastructure for unit tests and integration tests for data engineering pipelines using local services in docker-compose, which replaced existing manual tests.

YouTube Kids Web Team:

Developed primarily frontend features and infrastructure for the YT Kids web app, www.youtubekids.com, and YT parent tools app (primarily used in integrations), www.families.youtube.com. Developing efficient, robust, delightful UIs to support complex relationships between Google account types, permissions, and platforms across YT and Google and coordinating this across the stack was an area of significant complexity.

Developed individual appraiser allocation over the course of 2 months, which introduced new data models representing the appraiser entity, allocation status, and ranking in MongoDB and Elasticsearch.

Improved performance of API endpoints and batch processes by tuning database queries. This consisted of consolidating queries, limiting fields and preventing dereferencing, adding pagination, denormalizing data across collections, creating database indexes, and memoizing data.

Additional performance enhancements and fault tolerance were achieved through redesigning existing request-response flows to use a producer-consumer message queue pattern, which was implemented in Python and leveraged AWS SNS, SQS, and DynamoDB. These efforts saw latency reduced up to 90%.

Debugged concurrency issues with Datadog/Sentry monitors and added pessimistic locking.

Led agile stand-ups, retrospectives, and office hours, mentored rotational new hires, and conducted code reviews. Our team pioneered new processes that became standardized throughout the engineering teams, such as daily office hours, mocking in unit testing, and adding logging to increase platform visibility.

Designed the magnetic palettes product line for prototyping with 3D printing and further developed these designs for manufacturing with injection molding. Over 2,000 units of Palette 1.0 have been sold, with Palette 2.0 currently in production, and Palette 3.0 slated for release in 2021.