Verify

Drowning in misinformation? Verify uses AI to cross-check sources instantly, turning doubt into certainty. Stop the spread of fake news and empower your truth. Don't listen - Verify.

Comment

Inspiration

We spend hours consuming audio—podcasts, YouTube essays, and heated debates with friends. Unlike text, audio is fleeting and hard to fact-check on the fly. You can't "right-click" a spoken sentence to search for the truth. We wanted to build a tool that listens in the background and acts as an objective arbiter, catching misinformation the moment it’s spoken.

What it does

Verify is an AI-powered live fact-checker designed exclusively for audio. It monitors either your system audio (podcasts, videos, streams) or live conversations via your microphone. Verify keeps it simple and actionable: every claim is flagged as True, False, or Mixed. It’s essentially "Shazam for the truth."

How we built it

We prioritized a lean, high-performance architecture:

  • Frontend: Built with React and TypeScript.
  • Desktop Core: Developed using Electron to gain direct access to system audio interfaces.
  • AI Engine: Direct integration with the Gemini API.

Challenges we ran into

Our biggest hurdle was the Google Login and Authentication flow. To ensure a seamless user experience, we couldn’t ask users to manually generate and paste API keys. However, since the standard SDK doesn't natively support OAuth tokens for certain WebSocket modes, we had to manually implement the WebSocket communication from scratch. This allowed us to handle authentication and streaming stability ourselves without relying on a middle-man backend.

Accomplishments that we're proud of

We put user-centered design at the heart of our entire process. To us, "easy to use" meant more than just a clean UI, it required a frictionless login. By custom-building our own OAuth and WebSocket integration, we eliminated the need for manual API keys, making Verify truly accessible to everyone with a single click.

What we learned

We discovered a massive trade-off between context accuracy and latency. If the agent is too sensitive to context, the "statement extraction" phase takes too long. If it’s not sensitive enough, it fails to see the "big picture" and launches multiple separate searches for what should have been a single claim. Finding the "sweet spot" in our chunking logic was our most significant technical breakthrough.

What's next for Verify

Our immediate focus is Mobile Support to bring Verify into physical meetings and live events. We also plan to expand to other platforms and implement a "Session Summary" feature, allowing users to review all "False" or "Mixed" claims at the end of a long podcast episode

Share this project:

Updates