1 Created a RPC (Remote Procedure Call) Server which transfers data through JSON.2 Designed 10+ API endpoint protocols to expose part of the core business logic to the client side application.3 Separated the business logic into 2 layer where the lower level (service layer) is relatively more reusable than the high level (controller layer), so that the high level code is concise and the low level code is maintainable.4 Used Spring Boot technology to manage controllers and services as well as dependency injection.5 Implemented data access layer by leveraging JPA’s object relation mapping technology, which also involved ER data modeling of around 50 entities.6 Hand-wrote SQL to implement complex query to avoid JPA’s performance bottleneck.7 Use Apache Beam to batch process user’s log data into features that are useful for modeling click through rate in a distributed fashion.8 Design a gRPC server to execute long running jobs which is critical to remove memory bottleneck.

1 Built a module for identity verification using face image and other user features2 Implemented FaceNet like architecture to learn a vector representation of the face image3 Domain specific constraints were incorporated in an expert system to perform final filtering4 Used LFW dataset to pre-train the neural network, then a smaller internal dataset with location features was used for fine tuning5 Locations and mail sender features were used to schedule the training curriculum (triplet selection)6 Overall accuracy reaches more than 99% in a test dataset of 8,000 samples with hard mocked mailing information

1 Applied image processing to allow robots to more easily recognize the soccer ball under various lighting conditions 2 Implemented heuristics to control the velocity and acceleration which greatly improved the bot’s ball-keeping ability3 Using randomized decision rules for path planning to reduce Bayesian risk4 Won the championship