Cap-Check

Picture of the button
Picture of headband and a Cap
Picture of us slamming the button

Inspiration

Debates — whether political, academic, or casual — often get derailed by misinformation or outright lies. We wanted to create something bold, simple, and visual that anyone could use to call out falsehoods in real time. The idea of a big red button was both symbolic and fun, inspired by buzzer moments on game shows and the growing need for accessible fact-checking tools. We believe that misinformation is intolerable, and should be highlighted as such.

What it does

Cap-Check is an AI-powered truth buzzer for debates. It listens to the discussion, automatically partitions the audio stream into individual speakers, and keeps track of what each person has just said. When someone presses the big red button, the system instantly takes the most recent statement from that speaker and returns a verdict: True — 'Fact' supported by reliable sources Unclear — 'Sus' The statement can't be declared true or false False — 'Cap' likely misinformation or a lie This makes it possible to challenge questionable claims in real time, with both visual and digital feedback

How we built it

Frontend: Vite, React, TypeScript, Tailwind Backend: A Python server handling speech-to-text using assembly API, then running the text through a fine-tuned transformer model for semantic verification. Hardware: A physical “big red button” along with light up headbands wired to a Raspberry Pi pico.

Challenges we ran into

One major challenge was deciding between using one microphone or two microphones: With a single mic, we had to rely heavily on speaker diarization. This introduced problems when participants talked over one another, often resulting in incorrect partitions. With two mics, we reduced diarization complexity but ran into cross-talk issues, where both mics would still pick up the same speaker, creating confusion in the system. We also faced difficulties in connecting the backend and frontend. The frontend often failed to receive the correct data from the backend, and we struggled with reliably sending statements to Gemini for fact-checking. Debugging the communication between services under time pressure was a significant hurdle.

Accomplishments that we're proud of

Got a fully functioning prototype in under 48 hours. Managed to fully connect up backend and frontend Made a giant red button

What we learned

Learned how to better work with multiple audio inputs.

What's next for Cap-Check

Implement a gamified mode (3 lies before someone loses) Make into app. Add support for more debaters (for example, the primary elections have multi-person debates.