- Worked on the Permissions Science team to evaluate and improve automated testing for Cedar, a domain-specific langauge for writing authorization policies

- Evaluated and improved fuzz-testing input generators written in Rust to be more efficient and diverse at testing Cedar properties on the parser, validator, and authorizer.

- Found unique result of purely random generation to be stronger than coverage-guided fuzzing, ongoing work to publish results.

- Worked on Understanding team to improve traffic lights and emergency vehicle detection

- Built tools for traffic lights evaluation framework and improving triaging process

- Trained and deployed new flashing emergency vehicle detector for the Scene Understanding team

- Implemented a system that detects whether objects are occluding traffic light bulb detections using semantic segmentation, visual detections, and LIDAR camera projections as data sources, which was pushed to production on the car.

- Designed and implemented an ML pipeline for flashing traffic light detection, including scripts for data extraction, HTML / JS labeling and visualization tools, and experiments with various machine learning methods (Bayesian modeling, HMM, RNN).

- Architected and built a full ML pipeline for an RNN detector for flashing emergency vehicles, improving over the original model in precision and recall. Utilized CNNs and LSTMs for the model.

- Integrated the model into the Cruise emergency lights package using Tensorflow C++ API

- Implemented and deployed a Node JS/Express website that tracks soldier vests on Google Maps

- Added a Java API (using websockets) to integrate with the existing training app

- Presented to CEO and President of division