Signal.repair

Cut through the noise. Find what's real. Multi-agent AI that verifies any claim using live web search, 3-model consensus voting, and real-time source credibility analysis.

Comment

Inspiration

Every day we're bombarded with claims, AI-generated content, and information that may or may not be true. Fact-checking is slow, manual, and unreliable. We wanted to build something that could repair broken information signals automatically — using the live web, multiple AI models, and real evidence.

The domain signal.repair perfectly captured the idea: information is a signal, and when it's broken, noisy, or wrong, we repair it.

What it does

Signal.repair is a multi-agent AI system that verifies any claim, topic, URL, or AI-generated text in under 3 seconds.

Paste anything. Get the truth.

  • Scout Agent searches 50+ live sources via Tavily's real-time web APIs
  • Critic Agent scores credibility, assigns stances, finds contradictions
  • Consensus Engine runs 3 LLMs in parallel — majority wins, all disagree = DISPUTED
  • Signal Score — animated EKG oscilloscope showing 0-100 truth score
  • URL Analyzer — paste any article, get per-claim credibility scores
  • Claim Battle — two claims enter, evidence decides the winner
  • Trend Radar — real-time analysis of the world's most disputed topics
  • DNA of a Lie — visual timeline of how misinformation originates and spreads
  • Signal History — all repairs stored in Tower's lakehouse pipeline

How we built it

Backend: Python + FastAPI with Server-Sent Events for real-time streaming. Tavily for live web search. Groq for ultra-fast LLM inference across 3 models simultaneously. Tower for serverless pipeline orchestration and lakehouse storage.

Frontend: React + TypeScript with Framer Motion animations. Custom Canvas API for the EKG oscilloscope meter and animated source credibility network graph.

Architecture: User Input → Scout (Tavily) → Critic → Consensus Engine (3 LLMs) → Signal Score → Tower Lakehouse

Challenges

  • Building real-time SSE streaming with FastAPI's async generators
  • Getting 3 models to return consistent JSON for consensus voting
  • Building a force-directed graph on Canvas without D3
  • Tower deployment configuration and Towerfile format

What we learned

  • Ensemble LLM approaches produce significantly more reliable verdicts than single models
  • SSE is dramatically better than polling for real-time AI pipelines
  • Tower's serverless compute is genuinely powerful for AI pipeline scheduling

What's next

  • Nimble integration for deeper web extraction
  • Browser extension — right-click any text to repair it
  • Email alerts when Signal Watch detects topic changes
  • Enterprise API for newsrooms and research firms

Built With

Share this project:

Updates