Veriscope

We don’t decide what’s true — we show you the evidence.

Comment

🧠 Veriscope — Truth in Real Time

(wordplay on periscope) - Transcribes, Flags, Verifies

Hear it. Check it. Know it. A real-time Chrome extension that listens to any video and instantly shows supporting evidence for what’s being said — powered by AWS Bedrock, Amazon Transcribe, and Exa AI.

💡 Inspiration

In a world flooded with information — and misinformation — truth often lags behind speech. During live debates, news coverage, or roundtables, claims fly by faster than anyone can verify them.

We asked ourselves:

“What if truth could keep up with speech?”

That question became Veriscope — a tool that helps people see the evidence behind statements in real time, without bias or delay.

🔍 What it does

Veriscope is a Chrome extension that performs live fact aggregation for any online video.

While you watch, Veriscope:

  1. Listens to the audio in real time
  2. Transcribes speech using Amazon Transcribe
  3. Extracts factual claims using Claude 3 Haiku (AWS Bedrock)
  4. Finds supporting sources using Exa Search API
  5. Summarizes the evidence clearly and neutrally

On screen, a sleek sidebar displays each claim, supporting links, and a short AI-generated summary — updating dynamically as people speak.

🧠 Veriscope doesn’t decide what’s true or false. It simply shows you the evidence so you can decide.

⚙️ How we built it

  • Frontend:

    • Chrome extension that captures tab audio and streams it live via WebSocket
    • Real-time sidebar that displays claims, summaries, and sources

  • Backend (FastAPI):

    • Amazon Transcribe Streaming → converts audio to live text
    • AWS Bedrock (Claude 3 Haiku) → extracts factual claims & writes evidence summaries
    • Exa API (exa_py) → retrieves relevant web sources
    • WebSocket stream → pushes results back to the frontend continuously

Everything runs asynchronously for near real-time feedback — usually within 2–3 seconds of someone speaking.

🚧 Challenges we ran into

  • 🎙️ Audio streaming: Keeping latency low while handling PCM16 audio over WebSockets
  • 🧠 Claim extraction: Prompt-engineering Claude to output structured JSON without hallucinations
  • ⚡ Pipeline orchestration: Chaining AWS Transcribe → Bedrock → Exa → Bedrock again efficiently
  • 🧩 Summarization balance: Summarizing sources neutrally without assigning “truth” scores

🏆 Accomplishments that we're proud of

  • Built a fully working live audio fact-checker in just four hours 🔥
  • Integrated three complex APIs (AWS Bedrock, Transcribe, and Exa) seamlessly
  • Achieved low-latency real-time response (~2 seconds)
  • Designed a clean Chrome UI that feels intuitive and non-intrusive
  • Created an approach that promotes transparency over judgment

📚 What we learned

  • How to build streaming AI pipelines using AWS services
  • The art of prompt chaining and structured output for real-time systems
  • How Exa can make large language models more grounded by retrieving live, verifiable content
  • The importance of UX design in helping users trust AI outputs

🚀 What’s next for Veriscope

  • 🎧 Speaker identification: Attribute claims to specific participants
  • 🧮 Credibility scoring: Rank sources by reliability
  • 🗂️ Post-event summaries: Generate full recap fact-check reports
  • 🌐 Cross-platform support: Extend to podcasts, livestreams, and mobile apps
  • 💬 Community feedback: Let users rate and flag claims for transparency

🧩 Tech Stack

AWS Bedrock (Claude 3 Haiku) · Amazon Transcribe · Exa API · FastAPI + WebSockets · Chrome Extension (JavaScript) · Python

🎯 Tagline

“We don’t decide what’s true — we show you the evidence.”

Built With

Share this project:

Updates