Optimally perform fleet actions while facilitating touchless fleet transformations supporting Alphabet's capacity growth.

Designed and deployed a serverless engine on AWS to collect raw logs from Lambda API calls, EC2 systems and DynamoDB to form an organized, partitioned and SQL-queryable Data Lake and replicate DynamoDB tables hourly.

Used Glue to perform ET and data cataloging on semi-structured data, and extracted various identifiers to form mappings for correlation purpose.

Automated engine deployment using CloudFormation. The engine helped DevOps team on troubleshooting and monitoring with high scalability and low maintenance.

Proposed a differentiable function for propositional logic sentence to capture how well a probability distribution is to match the logic. Supplementing the function as a semantic loss to neural networks in order to improve probability of model outputs satisfying constraints.

Achieved (near-)state-of-the-art results on semi-supervised classification on MNIST & CIFAR-10, and increased ability of feedforward neural networks to predict structured objects, such as shortest paths and rankings, by 23% and 12% respectively.

Research paper accepted to ICML 2018.