At University at Buffalo, I built an automated profiling pipeline using Python, PyTorch, and GPU hardware counters to measure execution and energy use during deep learning training and inference. After finding that the measurement captured only 10% of BERT training energy, I redesigned it with execution hooks and isolated module replay, increasing coverage to 94.6%.
At Tala, I designed a shared Java Core + Adapter library and extended services to generate customer lifecycle events for attribution. I also redesigned shared authentication state to address login failures affecting 100+ daily users, and migrated role-based access control across 15+ services.
I built and actively develop DealFlow, a lead and employee management platform used by a working car dealership, and am building Mergewright, a multi-agent coding system that turns GitHub issues into validated pull requests. Earlier, at SenSight Technologies, I migrated over 100 GB of data from MySQL to BigQuery and designed Looker dashboards.

