Guide and mentor software development team to apply machine learning, geospatial, and photovoltaic modeling solutions to design solar panel systems for residential properties.

Develop and deploy machine learning, geospatial, and photovoltaic modeling solutions to design solar panel systems on residential properties.

* Collaboratively designed, implemented, and deployed a distributed system in AWS to identify and classify elements in thousands of geospatial photographs per hour* Increased deployment stability rates from 50% to 95% and decreased deployment time by 50% by creating streamlined deployment procedures* Delivered a 50% cut in AWS data transfer costs by guiding the transformation to a smaller image format in relaying images between clients and microservices* Eliminated 20% processing waste by fixing the image tiling algorithm to not duplicate edge tiles* Converted neural network weights from legacy framework caffe to pytorch with no net difference* Orchestrated the addition of hundreds of images to the test set for roofing material classification* Balanced building detection quality and speed by tuning image scale and tiling parameters* Amortized docker build time by 67% using a multistage dependencies-then-app pattern* Standardized code quality and style across dozens of repositories by adopting the pre-commit tool* Increased development time by 67% by streamlining workstation configuration from 3 days to 1

Implemented architectural sketch functionality features as part of the Xactimate property insurance claims estimation software

Master's thesis research* Studied machine-assisted annotation for suggesting Semitic root words for Syriac and Arabic text* Adapted training for a Markov model to be done incrementally on partially annotated sentences* Simulated text annotation in reading order and word frequency order and demonstrated the machine-learned model had faster and more steady performance gains than the baseline algorithm

* Constructed weighted graph of similar names from 500,000 paired name records; compared to existing name lists, the Levenshtein-based string edit distance and feature vector dot product alignment methods achieve 97% recall and 4.5% and 5.5% fall-out respectively* Crafted a modular application infrastructure in Java to reuse operations in multiple applications and keep command line interfaces consistent across applications

* Led development of a data governance web application to manage data sharing agreements* Stabilized 15 of 23 pages by eliminating null pointer exceptions and validating client input* Engineered essential features enableing 60% of new agreements to be completed in the application* Reduced code duplication and began incremental refactoring towards a more maintainable and extensible design

Implemented three features for the X5 release of the Chief Architect CAD software, which all involved constructing menus and dialogs, representing objects in 2D graphics, facilitating interactive editing, and extending intra-file transfer of default settings