Verity

Inspiration

A lot of bad decisions do not come from bad intentions. They come from incomplete information.

When people hear a confident claim in a video, podcast, or debate, they often accept it, reject it, or repeat it before checking the evidence. Over time, that can harden into polarized opinions and lead individuals, teams, and communities in the wrong direction.

We built Verity because we wanted a simpler way to pause that cycle.

Our goal was not to create an AI that tells people what to think. It was to create a calm, transparent guide that helps people see what is supported, what is misleading, and what still needs more context.

What it does

Verity is a real-time fact-checking companion for YouTube.

While a video plays, Verity listens for factual claims, checks them against trusted documents and web sources, and displays a clear verdict with citations directly on the video. The same result can also be sent to a paired iPhone.

For our demo, Verity evaluates the claim:

“Electric vehicles produce no carbon emissions.”

Rather than giving a simple true-or-false answer, Verity explains the nuance: electric vehicles have no tailpipe emissions, but emissions can still come from electricity generation, battery production, and the broader manufacturing lifecycle.

That context is often what helps a conversation move forward.

Why it matters

Verity is designed to support mental clarity, better decisions, and healthier conversations.

During a debate, it can help people separate facts from opinions. During a meeting, it can help a team verify an important claim before acting on it. During everyday media consumption, it can help someone slow down before sharing something misleading.

The goal is not to “win” an argument. The goal is to create enough shared understanding for people to make better decisions together.

How we built it

A Chrome extension captures audio from the YouTube tab and streams it to a FastAPI backend. Deepgram transcribes the conversation and identifies speakers. AI detects checkable factual claims, then a DigitalOcean Gradient agent gathers evidence from a trusted PDF knowledge base, using web search only when additional evidence is needed.

A reasoning model drafts the verdict, but the AI does not get the final word.

Deterministic software verifies every citation, checks that quoted evidence actually exists, requires multiple credible sources, and returns Insufficient Evidence when the result cannot be supported.

The final verdict is stored in PostgreSQL, shown in a React overlay on YouTube, and sent through Web Push to an installable iPhone PWA.

Challenges we faced

The biggest challenge was balancing speed with trust.

A real-time system needs to respond quickly, but a fact-check is only useful when people can understand where the answer came from. We had to coordinate streaming audio, transcription, claim detection, evidence retrieval, citation validation, browser updates, and phone notifications without allowing weak or invented evidence to reach the user.

We also designed every AI provider with a transparent fallback, so the full experience can still run if an external service fails during the demo.

What we learned

We learned that AI is most valuable here as a guide, not an authority.

Models are good at detecting claims, finding patterns, and explaining evidence. But trust comes from the surrounding architecture: visible sources, deterministic validation, uncertainty, and the willingness to say, “There is not enough evidence.”

We also learned that fact-checking is rarely just about labeling something true or false. The most valuable answer is often the missing context that helps two people understand each other.

What’s next

Next, we want to expand Verity beyond one YouTube demo into a tool for live debates, meetings, classrooms, interviews, and public discussions.

Our larger vision is a world where people can disagree without becoming disconnected from the facts, and where technology helps us find common ground instead of pushing us further apart.