signal.repair

• 

  signal repair

Inspiration

Misinformation spreads faster than people can fact-check it. Caregivers, nurses, journalists, and family members often see viral health claims and don’t know if they’re true. We wanted a tool that doesn’t just translate text — it repairs broken information signals by checking claims against live web evidence. The domain signal.repair fits that idea: when the signal is wrong, we repair it with cited sources.

What it does

signal.repair is an agentic fact-checking web app. Users paste a health or news claim (e.g. “Drinking bleach cures COVID-19”). The system:

Extracts the core claim Researches live web sources via Nimble Search API Scores source trust Returns a verdict — Supported, Mixed, or Disputed — with citations and a full markdown report Built with a Gradio UI so anyone can use it in a browser. Demo video: https://youtu.be/ONtJ7mwh0xE

How we built it

LangGraph — 3-agent pipeline: Extract → Research → Verdict Nimble API — live web search for every verification MCP (FastMCP) — typed tools (search_allowlisted_sources, verify_claim) with domain guardrails Source scorer — trust weighting by domain tier Gradio — full-width UI for claim input and results Python stack: httpx, scikit-learn Architecture adapted from a prior multi-agent + MCP project (SIFT-Analyst), rebuilt for misinformation / fact-checking instead of forensics.

GitHub: https://github.com/prakash023-hub/signal-repair

Challenges we ran into

Nimble API tiers — extra parameters (include_domains, focus) returned 403 on hackathon keys; fixed by using minimal search payload and scoring trust in-app UI in Gradio dark theme — footer text was white-on-light; fixed with explicit colors Layout — textbox was squeezed beside the button; moved to full-width stacked layout for better demo UX API key setup — documented .env flow so judges can run locally with their own Nimble key

Accomplishments that we're proud of

End-to-end agentic app with live Nimble web search (not mock data) Working verdicts on real demo claims (e.g. bleach/COVID → DISPUTED with PubMed/CIDRAP sources) MCP guardrails — architectural allowlist, not prompt-only Clear domain story: signal.repair for DeveloperWeek Nimble + name.com Domain Roulette tracks Open-source repo, README, demo video, and runnable Gradio app Built solo by Prakash Raj for DeveloperWeek New York 2026.

What we learned

How to integrate Nimble Search API into a LangGraph multi-agent workflow Sponsor API constraints (trial vs enterprise parameters) and how to design around them MCP typed tools as guardrails for agentic apps Building for general tech judges (clear problem, live demo, startup feasibility) vs niche domains

What's next for signal.repair

Deeper Nimble search (fast / deep depth) when enterprise API access is available Browser extension — fact-check highlighted text on any page Real-time monitoring — alert on trending false health claims More languages — Tamil/Hindi claim support Mobile PWA with camera/OCR for claims on printed material Deploy permanent demo on Hugging Face Spaces

Built With

• api
• fastmcp
• gradio
• httpx
• langgraph
• mcp
• nimble
• python