I explore problems in complex systems, build computational and experimental tools to understand them, and turn ideas into open-source software products.
I am interested in making the assumptions behind a computational decision visible, so that a prediction can be examined rather than simply accepted.
PocketStop uses early molecular motion to flag unpromising simulation runs. Calibrated decisions help preserve runs that may still pass the screening endpoint.
A retrospective replay of completed simulations, with held-out proteins excluded from model fitting and threshold calibration. The policy has not yet been tested on new screens.
As sole author, I defined the problem, ran the simulations and machine-learning analyses, and released the software.
CookieGuard AI helps build trust in AI browser agents by independently checking how their login cookies are protected. It also supports browser tests, CI, and interactive Chrome scans. 1,852 reported downloads across 6 npm and PyPI packages (through Nov 5, 2026; 1 package awaiting its first report; includes recorded snapshot).
I built the Chrome extension and expanded the project into tools for browser agents, automated tests, and CI.
CookieGuard AI · Chrome extension · Developer toolkit
AirWater AI uses climate and material data to screen sites and operating schedules. It flags cases with weak evidence.
Phoenix in July with a 3 L/day target: an hour-by-hour capture-and-release plan and the top-ranked material, Al-Fumarate, at an estimated 3.47–4.50 L/day-equivalent before device losses. For the same site in January, the application answers Do not deploy.
I built it as a functional proof of concept: the isotherm dataset, the uptake model, the ranking and scheduling logic, and the web app.
AirWater AI · Public web application
Recognition: International Youth AI Competition 2026 (Fall) — Gold Award
BenefitStep reads pay stubs, bills and receipts on the user's own computer, links each detail to its source, and fills in unsigned drafts of California's official forms.
The extension's first and last steps, from choosing an application to generating the draft forms. In a test of its document reader on 26 fictional files written for the project, 66 of 68 filled-in values matched the document.
I researched the problem, designed the workflow, and directed AI-assisted implementation and testing.