Verdict

Every hackathon is a courtroom. Verdict is the agent that's read every case file before you walk in. Verdict is an autonomous Gemini-powered agent system with two modes: SCOUT — Point it at any hackathon URL, or feed it submissions directly. The agent enriches each project with AI-written descriptions and extracted links (GitHub, demo, video), clusters them by theme, and produces a competitive intelligence report: who's building what, what gaps no one is filling, and which dark-horse project is positioned to win. JUDGMENT — Bring your own judging rubric in plain English ("novelty 40%, technical depth 30%, real-world impact 30%") and a batch of submissions from any source: scrape a URL, upload a file (CSV, JSON, Markdown, PDF, TXT), or paste them manually. Verdict scores every submission against your rubric and delivers a ranked leaderboard with per-criterion scoring and judicial verdict lines. Both modes stream the agent's reasoning live via Server-Sent Events to a courtroom-themed interface — you watch the agent think, fetch, deliberate, and rule in real time. The core is a function-calling agent loop on Gemini. The reasoning model (Pro tier) handles planning, clustering, and scoring. The fast model (Flash tier) handles high-volume tool steps: link extraction, description writing, file parsing, and rule parsing. Users can swap either model from the UI. Every ruling gets a permanent shareable URL — pass verdicts around like screenshots. Tech stack: Python · Flask · Gemini (function calling, dual-model) · Server-Sent Events · BeautifulSoup + Playwright · pypdf · pandas · gunicorn + gevent. Designed to deploy on Vultr Ubuntu 24.04. Built solo for the AI Agent Olympics at Milan AI Week 2026.