Verify News
AI | BUILT | by Roberto Salodini
https://robertosalodini.com/work/verify-news

An AI news checker that explains its verdict, and goes where people read: a bookmarklet, the phone share sheet and YouTube links. Live and open source.

Role: Designer and developer
Stack: Next.js (https://nextjs.org), React (https://react.dev), Tailwind (https://tailwindcss.com), Llama 3 (https://www.llama.com), Brave Search API (https://brave.com/search/api/), built with Cursor (https://cursor.com) and Gemini Pro
Live: news-verifier-mu.vercel.app (https://news-verifier-mu.vercel.app) · GitHub (https://github.com/robrego/news-verifier)

THE PROBLEM
A fake or real label tells you nothing about why, and people meet doubtful claims inside feeds and long posts, not on a fact-checking website. I wanted to move past a static Figma prototype and build something that works the way a reader does.

CALL 1: EXPLAIN THE VERDICT
A Next.js backend scrapes the article and checks its claims, with Llama 3 and the Brave Search API bringing in real-time sources. To curb hallucinations I added strict rules: the system extracts each article's publication date for temporal context.

CALL 2: START FROM ONE BOX
The News view is one box: paste a headline, a claim or a link, with claims to try. The app returns a trust score and a forensic analysis, so the reader sees what was checked.

CALL 3: DESIGN FOR LIMITS
The interface shifts into a masonry grid depending on how much evidence comes back, so dense results stay readable. Technical limits became features: a custom state handles API limits, and pasting a new link clears the screen with zero clicks.

CALL 4: MEET READERS WHERE THEY READ
Copying a link into a website is friction most people will not repeat. A Tools page offers two shortcuts so nobody has to paste by hand:

- On a computer: a bookmarklet, "Verify this page", that checks the page you are on, or just the text you selected.
- On a phone: add the app to the home screen on Android and it appears in the share sheet of other apps. iPhone does not support this yet, so the page offers a fallback: copy the link or upload a screenshot.

CALL 5: BEYOND TEXT
A Videos section takes a pasted YouTube link and runs a deep dive on it, because many doubtful claims arrive as video.

RESULT
- A live, full-stack RAG product, shipped by one person and open source on GitHub
- Checks articles, headlines, selected text, screenshots and YouTube links
- Works where people read: bookmarklet on desktop, share sheet on Android
- Light and dark themes and a custom state for API limits

WHAT I LEARNED
The constraints were the design brief. API limits, platform limits and the model's own mistakes each became a visible, honest part of the interface, and meeting readers inside their feeds mattered more than polishing the paste box.