Inspiration
I kept catching AI chatbots confidently stating things that just weren't true — so I figured, why not make the bot check its own homework before showing it to me?
What it does
It answers your question, then actually goes and verifies that answer against real sources, giving you a confidence score, warnings, and the sources it used — instead of just trusting the first response.
How I built it
FastAPI backend with Gemini generating and verifying answers, Tavily for web search, and a curated list of ~1,000 trusted domains to score source credibility — all wrapped in a React chat UI with multiple chats, dark mode, and citations you can click into.
Challenges I ran into
Getting the confidence score to mean something real (not just a random number) took a few iterations, and I caught a sneaky bug where the verifier was only ever looking at the first search result instead of all of them.
Accomplishments that I'm proud of
The verification pipeline actually flags when sources disagree or are outdated, and the confidence score comes with real reasoning attached — it's not just a number, it explains itself.
What I learned
Trusting an LLM's first answer is easy — building a system that makes it prove itself is a whole different (and much more interesting) problem.
What's next for AI Fact Shield
User accounts with synced chat history, claim-by-claim breakdowns instead of one score per answer, and letting people paste a URL or screenshot to verify instead of just typing a question.
Log in or sign up for Devpost to join the conversation.