Example of live fact checking from scrolling on a social media feed.
Example of a series of statements made, and verif.ai determining if they are true.
Summary page, synthesized after conversations end.
Inspiration: ground news, news checking AI agents. Increasing amount of misinformation these days with increased social media use and volume of media made with AI.
What it does: takes live audio, splits between multiple speakers, and performs live truth verification through an ensemble of models and pulling real news evidence. Has transcription, classification, and verification modules. Reports truth of statements and summarizes the accuracy of speakers throughout a conversation.
How we built it: ElevenLabs, OpenAI, and Perplexity API to transcribe audio, determine relevance and route segments, and then verify or dispute information. Frontend web interface to track transcripts, gives truth score over the duration of conversation, grades speakers at the end.
Challenges we ran into: accurate speaker separation during transcription, having low latency between speaking and verification.
Accomplishments that we're proud of: Accuracy of fact checking, cleanliness of UI.
What we learned: How to use live AI audio agents, process and chunk segments for optimal latency and classification accuracy.
What's next for verif.ai: We want to expand the mediums by which you can use fact checking. We want better integrations with everyday information sources such as social media, articles, news streams, and daily conversations. We want to develop a mobile app as well as a more agentic framework where a single user can fact check more sources at once.
Log in or sign up for Devpost to join the conversation.