Implemented convolutional neural nets for checkbox localization and evaluation using single shot detector (SSD) architecture from partially automated data collection through hyper-parameter optimization for a validation set accuracy of 94%.

Created a distributed workflow server architecture to process prioritized workflow concurrently and scaled deployment to more than 150 servers processing millions of pages per day.

Migrated and containerized workflow system and other application functions into micro-services enabling runtime scaling of processing servers with AWS managing infrastructure.

Diagnosed and cleaned data used in support vector machines (SVMs) to classify document pages into thousands of classifications with an accuracy of up to 93%.

Optimized processes utilizing multithreading balancing CPU usage with I/O load cutting processing time.

Automated the setup for the natural language processing module saving an average of 50 work hours per installation.

Designed and implemented logic for automatic property legal description extraction within an existing auto-indexing module.

Designed a multithreaded bulk processor that decreased compute time by varying degrees (2x-10x) and improved resilience for CPU intensive, commonly error-prone processes.

Optimized database queries used in hardcoded county financial reports to decrease run time by up to a factor of 40, in one case decreasing running duration from 5 days to 3 hours.

Presented Treasurer “Advanced Topics” to client power users at Tyler User Conference.

Advised customer support team with technical assistance connecting directly with clients as needed.

Developed and maintained Recorder software packages as part of a 12-person team and both Treasurer and Assessor software packages as part of a 9-person team.