clarity-bot

AI-powered misinformation fact-checker using OpenClaw

Comment

Inspiration

Misinformation spreads faster than corrections. We wanted to make fact-checking as easy as submitting a form. Paste any claim, get a scored verdict with sources in seconds.

What it does

ClarityBot takes any claim and runs it through a 5-step AI pipeline: assertion extraction, web source retrieval, cross-referencing with credibility labels, credibility scoring (1–10), and verdict labeling. Every step streams live to a real-time dashboard terminal via Server-Sent Events, so users can watch the agent think.

How we built it

  • NemoClaw (NVIDIA, released March 16 2026): sandboxes the OpenClaw agent with a deny-all network policy via OpenShell. Every outbound call is governed — a prompt injection attack cannot exfiltrate data.
  • OpenClaw: agent framework with 5 custom Python skills (extractor, searcher, crossref, scorer, emitter).
  • Gemini 2.5 Flash: claim extraction, cross-referencing, and scoring via the REST API with retry + backoff for rate limits.
  • Serper.dev: Google search results for source retrieval — 3 assertions × 3 results, deduplicated by URL.
  • Source credibility filter: trusted-source registry covering Space/NASA, Vaccines/Health, and Historical facts. Scores are capped based on source quality (primary authority → no cap, no high sources → capped at 6, only low sources → capped at 4).
  • FastAPI (Python): REST backend + Server-Sent Events for live streaming. All Supabase calls wrapped in asyncio.to_thread() to avoid blocking the event loop.
  • React + Vite + TypeScript + Tailwind: custom dark-mode frontend with animated pipeline steps, donut/bar charts (plain SVG, no charting libraries), and a terminal-style agent logs panel.
  • Supabase: hosted Postgres for claims, logs, and trend data.
  • Vercel: deployed as a single project — FastAPI serverless backend
    • pre-built React frontend served via CDN. DB-backed SSE polling replaces in-memory queues for serverless compatibility. ## Challenges we ran into
  • Configuring NemoClaw's OpenShell policy to allow only necessary endpoints while blocking everything else by default.
  • Syncing the SSE stream with the frontend's animated pipeline step UI — the queue must be created before the background task starts, or the first events are lost.
  • Wrapping synchronous Supabase client calls in asyncio.to_thread() to prevent blocking the FastAPI event loop.
  • Tuning the source credibility filter so that scores accurately reflect evidence quality across different topic domains.
  • Deploying to Vercel's serverless platform required replacing in-memory SSE queues with DB-backed polling and splitting the check/execute flow for stateless function compatibility. ## Accomplishments that we're proud of
  • The live pipeline animation: watching 5 steps light up in real time while the agent searches, cross-references, and scores.
  • The source credibility system: not all sources are equal, and our scoring reflects that with hard caps based on evidence quality.
  • The NemoClaw sandbox: a real security story, not just a checkbox. ## What we learned NemoClaw was released 12 days before this hackathon. Learning its blueprint lifecycle and sandbox policy system was the most technically interesting part of the build. We also learned how powerful SSE is for real-time agent observability, no WebSocket complexity needed. ## What's next for clarity-bot
  • Telegram bot via NemoClaw's built-in bridge
  • Browser extension to check claims inline on social media
  • Multi-language support for non-English claims
  • Batch checking mode for journalists processing multiple claims

Built With

Share this project:

Updates