Hamzah Mashabi

<p>• Built the alliance's project-evaluation agent: Gemini reads each team's PDF once, three judge runs score it against a nine-criterion anchored rubric out of 100, and code does all the arithmetic, quote checks and report building, so no model sits between the results and the page. • Integrated it into the alliance's website (Next.js, Prisma, MariaDB) across six pull requests: a background worker that runs every minute, saves every stage so an interrupted run resumes, and retries after 10 and then 30 minutes; teams receive a private report link by email and can resubmit up to three versions to compare their levels. • Built the guardrails: phones and emails stripped and member names anonymised before any model call, defences against prompt injection in uploaded files with flagged reports withheld for human review, and an explicit consent step before AI evaluation. • Measured the evaluator before trusting it: a two-line rubric rewording lifted one project from 70 to 81 on the same reading, so I reverted it and made side-by-side comparison on real submissions mandatory for any judge-facing text; across 8 submissions, 7 stayed within 2 points between readings.</p>