VeriTounes

Your AI watchdog for fighting scams and fake news in Tunisia.

Comment

Inspiration

In Tunisia, the rise of digital communication has brought with it an explosion of scams, fake news, and misleading messages — often spreading through WhatsApp voice notes, Facebook posts, and image-based messages. Many citizens, especially elders and those in rural areas, are vulnerable due to the lack of accessible verification tools in Tunisian dialect. We were inspired to build VeriTounes — an AI-powered solution that understands how Tunisians communicate and helps verify information in a way that’s local, simple, and trustworthy.

What it does

VeriTounes is a multi-input AI assistant that helps users:

  • Detect scam messages or fake news from text, images, or voice notes.
  • Analyze content for red flags using transformer models and fact-checking APIs.
  • Explain its findings in Tunisian dialect or Modern Standard Arabic for better accessibility.
  • Score content with a trustworthiness index (0–100) with traffic light indicators: 🔴 red (likely scam), 🟡 yellow (needs attention), 🟢 green (safe).

How we built it

  1. Dataset Curation: Collected real scam texts, voice notes, and images circulating in Tunisia. Annotated them manually for supervised training.

  2. Modeling: Fine-tuned XLM-RoBERTa for text classification (scam vs legitimate). Used Whisper for voice-to-text conversion. Integrated Tesseract OCR to extract text from screenshots.

  3. Frontend & UX: Built a responsive web app using React and Tailwind CSS. Developed a Telegram bot for quick access via mobile.

  4. Explainability & Language Adaptation: Integrated NLLB (No Language Left Behind) for dialect translation. Used rule-based templates to explain AI decisions in friendly language.

Challenges we ran into

  • Dialect Complexity: Tunisian Arabic lacks standardized spelling and grammar, making text processing difficult.
  • Voice Data Limitations: Voice messages often contain background noise, slang, or code-switching (Arabic-French).
  • Real-Time Accuracy: Balancing speed and accuracy when processing multi-modal inputs.
  • User Trust: Ensuring users understand and trust the AI output without it feeling too “robotic” or technical.

Accomplishments that we're proud of

  • Successfully built a trilingual AI assistant (Arabic, French, Tunisian dialect).
  • Enabled multi-input detection (voice, text, image) in a single seamless interface.
  • Created a public Telegram bot that can be used even by non-tech-savvy users.
  • Built a trust score system with visual feedback that helps users make better decisions.

What we learned

  • AI needs to adapt to local culture and language to be truly impactful.
  • Data quality matters more than quantity — small, well-annotated local datasets performed better than general ones.
  • Explainability = Adoption — People will use AI tools only if they feel the outputs are clear and relevant.
  • The power of team collaboration and user feedback in rapid prototyping.

What's next for VeriTounes

  • Deploy as a public web app & mobile-friendly site with real-time inference.
  • Partner with Tunisian media literacy organizations to train users.
  • Expand detection capabilities to video content (e.g., TikTok clips).
  • Localize further using community-contributed scam reports to improve accuracy.
  • Apply for government and NGO support to deploy in rural areas and schools.

Built With

Share this project:

Updates