A fake news detector that actually works. Submit a headline, URL, or screenshot. Get a verdict (REAL/FAKE/UNCERTAIN) with confidence scores and sources that aren't just hallucinated by an AI.

Live demo: https://truthgazette.vercel.app — give it a whirl!

Easter egg: Press Shift+L on the site to toggle "lights-out" flashlight mode. It's dumb, but cool.

• Text: Paste a claim → AI analyzes it + Google search
• URL: Drop a link → We fetch it, parse it, verify the sources it cites
• Image: Upload a screenshot → OCR extracts text, we fact-check what's in it
• Output: Verdict + confidence (65-95%) + why we think what we think + actual verified sources

The whole thing is built to prevent the AI from just making stuff up. Every URL it cites gets fetched and validated. If a source is dead, we check the Wayback Machine. If the excerpt doesn't exist on the page, we flag it.

npm install
npm start

Open http://localhost:3000. It's a newspaper-style interface (because we're committing to the bit).

Quick demo tips:

• Paste a headline or drop a URL, hit "INVESTIGATE NOW".
• Upload a screenshot for OCR; cropping tight around text helps accuracy.
• Try the live site first: https://truthgazette.vercel.app

• Frontend: Plain HTML + JavaScript (no framework flex needed)
• Backend: Node.js serverless function on Vercel
• AI: Google Gemini 2.5-flash with Google Search grounding
• OCR: Google Vision API (primary) + Tesseract.js (fallback)
• Rate Limiting: Upstash Redis (prod) or in-memory Map (local)
• Source Verification: Custom function that actually fetches URLs and checks the HTML

GEN_API_KEY=your-gemini-api-key
GEN_MODEL=gemini-2.5-flash
UPSTASH_REDIS_REST_URL=...  # optional for production
UPSTASH_REDIS_REST_TOKEN=...
RATE_LIMIT_PER_MIN=20
DAILY_QUOTA=200

Don't have Upstash? Cool, it'll just cache in memory locally.

.
├── index.html           # UI + all client-side logic
├── api/investigate.js   # Backend that does the real work
├── package.json         # Dependencies
└── favicon.svg          # A tiny newspaper

• Web Archive snapshots are sometimes incomplete or outdated
• Date extraction from HTML uses regex, not NLP magic
• We block private IPs (no localhost scanning)
• Max 10MB images (Google Vision gets cranky)
• If the model is having a bad day and returns "no content", we retry once

• 20 requests/minute per session
• 200/day per session
• Cached for 1 hour so you don't hit the API 500 times with the same query

Capstone project by Ishant. Built to actually work instead of just being a fancy demo.

MIT. Go nuts.