An AI-powered claim-level fact verification for articles, news, and images, built to fight misinformation in real time.

---

How it works

Paste an article URL or text. Claude Haiku breaks it into atomic, verifiable claims. Each claim is cross-referenced in parallel against Google Fact Check, Wikipedia, GDELT, and ClaimBuster. A DeBERTa NLI model scores every evidence passage for entailment or contradiction, then Claude Sonnet synthesises a per-claim verdict with confidence and cited reasoning. Results stream to the browser via SSE as they arrive.

For images (screenshots, WhatsApp forwards, viral posters), Tesseract OCR extracts the text first, falling back to Claude Vision when confidence is low. The same evidence and synthesis pipeline then runs on the extracted claims, alongside nine manipulation signal detectors that flag urgency language, fake authority claims, missing attribution, and similar patterns.

---

Features

• Claim-level granularity - verdicts per statement, not per domain
• 4-band verdict system: Supported · Refuted · Mixed · Insufficient Evidence
• NLI contradiction detection using DeBERTa-v3 across all evidence passages
• Web search fallback for claims too fresh or niche for structured sources
• Image fact-checking: OCR → claim extraction → evidence → verdict
• Real-time streaming via Server-Sent Events - no polling
• Every verdict cites the passages it reasoned from
• ~$0.006 per article check on the standard path

---

Tech Stack

Layer	Technology
Backend	Python 3.12, FastAPI, Pydantic v2, httpx
AI Models	Claude Haiku 4.5 (extraction), Claude Sonnet 4.6 (synthesis), Claude Vision (OCR fallback)
NLI	DeBERTa-v3-base-mnli-fever-anli (HuggingFace Transformers)
Evidence Sources	Google Fact Check API, Wikipedia, GDELT, ClaimBuster
OCR	Tesseract (primary), Claude Vision (fallback)
Streaming	Server-Sent Events (sse-starlette)
Frontend	Vite, React 18, TypeScript, Tailwind CSS v3, Framer Motion

---

Getting Started

Prerequisites

• Python 3.12+
• Node.js 20+
• An Anthropic API key
• Tesseract OCR binary (optional - falls back to Claude Vision if absent)

# Ubuntu / Debian / WSL
sudo apt-get install -y tesseract-ocr

# macOS
brew install tesseract

# Windows - installer at https://github.com/UB-Mannheim/tesseract/wiki

Installation

Backend

cd backend
python -m venv .venv
source .venv/bin/activate        # Windows: .venv\Scripts\Activate.ps1
pip install -r requirements.txt
cp .env.example .env             # then edit .env
uvicorn main:app --host 127.0.0.1 --port 8000 --reload

Frontend

cd frontend
npm install
# create frontend/.env with: VITE_API_URL=http://127.0.0.1:8000
npm run dev

Open http://localhost:5173.

Environment Variables

Variable	Required	Description
ANTHROPIC_API_KEY	✅	Anthropic API key
ANTHROPIC_EXTRACT_MODEL	no	Defaults to claude-haiku-4-5-20251001
ANTHROPIC_SYNTHESIS_MODEL	no	Defaults to claude-sonnet-4-6
GOOGLE_FC_API_KEY	no	Enables Google Fact Check source
CLAIMBUSTER_API_KEY	no	Enables ClaimBuster source
NLI_ENABLED	no	Set to false to disable DeBERTa (faster, less accurate)
FRONTEND_ORIGIN	no	CORS origin, defaults to http://localhost:5173

---

API Reference

Article pipeline

POST /check

{ "input": "https://example.com/article  OR  pasted article text" }

Returns { "check_id": "<uuid>" }

GET /check/{check_id}/stream - SSE

Event	Payload
claim_extracted	{ id, text, entities }
nli_scored	{ claim_id, supporting, contradicting, verdict_band }
source_results	{ source_health }
web_search_triggered	{ reason }
web_search_complete	{ results_count, sources }
verdict	{ article_verdict, disclaimer }
done	{ status }
error	{ message }

Image pipeline

POST /image-check - multipart, field file, max 10 MB
Returns { "check_id": "<uuid>" }

GET /image-check/{check_id}/stream - SSE

Event	Payload
ocr_complete	{ text, confidence, language, method }
claim_found	{ claim: { id, text } }
signals_ready	{ signals, signal_score }
verdict	{ verdict_label, band, confidence, reasoning, manipulation_signals }
done	{ status }
error	{ message }

---

Contributing

Open an issue before submitting a PR. Bug fixes and well-scoped features are welcome.

---

License

MIT