At ECWoC & SWoC Programs, I designed and implemented graph visualization and Graphviz DOT export for an open-source C++ graph library, addressing a usability gap for new users. I worked with the project maintainer through design discussions and code review before the changes were merged upstream.
I built CleanDrop, a local-first download manager with a Chrome extension and Windows background service. It uses llama.cpp to generate filenames without cloud services, and I redesigned its processes to support asynchronous work safely.
For CleanDrop, I used SQLite WAL mode to cut mean insert latency from 45ms to 3ms, and added identity-based file tracking so deletion remains safe after files move or are renamed. I also added a local MCP server with four tools for querying and cancelling pending deletions, backed by least-privilege safeguards and 29 tests.
I built Fido, a real-time synchronized watch platform that maintained about 80ms average drift from the server timeline; load testing validated a concurrency fix across 9,400 connections with zero failures. In a separate star-field localization project, I built an evaluation pipeline and reduced deep-learning model error from about 80° to 7.8°.
