ثقة Thiqa

AI-Powered Seller Trust Verification Platform for the Algerian Digital Economy

Thiqa (ثقة — Arabic for trust) helps Algerian online buyers assess the trustworthiness of social media sellers before completing a purchase. Paste any Facebook, Instagram, or TikTok seller URL and get an instant AI-generated trust score with a verdict in Algerian Darija.

---

✨ Features

• 🔍 Seller Trust Score (0–100) — Multi-signal Gradient Boosting + rule-based blend
• 🗣️ Arabic/Darija Sentiment Analysis — Fine-tuned MARBERTv2 on Algerian e-commerce text
• 🖼️ AI Image Detector — Sightengine API + Groq LLaVA fallback
• 📸 Screenshot OCR Pipeline — EasyOCR extracts text from conversation screenshots
• 🤖 Scam Text Analyzer — LLaMA 3.3 70B via Groq with few-shot Algerian scam examples
• 📋 Community Reports & Reviews — Credibility-scored report system
• 🚫 Blacklist — Paginated, searchable blacklist of reported scammers
• 🔐 JWT Authentication — Stateless auth with bcrypt password hashing

---

🧠 AI Pipeline

User Input (URL / Text / Image)
        ↓
Platform Detection & Scraping (Apify)
        ↓
┌───────────────────────────────────────┐
│  EasyOCR  │ MARBERTv2 │ Sightengine  │
│ Screenshot│ Sentiment  │ AI Image Det │
└───────────────────────────────────────┘
        ↓
Gradient Boosting + Rule-Based Blend
        ↓
Trust Score (0–100) + Darija Verdict

Trust Score Verdict Thresholds

Score	Color	Label
80–100	🟢 Dark Green	Very Trustworthy
65–79	🟩 Green	Generally Trustworthy
50–64	🟨 Yellow	Some Doubts
40–49	🟧 Orange	Concerning Signs
0–39	🔴 Red	Avoid

---

🏗️ Tech Stack

Layer	Technology
Frontend	React 18 + Vite
Backend	FastAPI + Uvicorn (Python 3.11+)
Database	PostgreSQL 15 + SQLAlchemy + Alembic
LLM	Groq API (LLaMA 3.3 70B)
Sentiment	MARBERTv2 (fine-tuned) + CAMeL BERT fallback
OCR	EasyOCR (Arabic + English)
ML	scikit-learn GradientBoostingClassifier
Image AI	Sightengine + Groq LLaVA fallback
Scraping	Apify (Facebook / Instagram / TikTok)
Storage	Cloudinary (report screenshots)
Auth	JWT (python-jose) + bcrypt

---

🚀 Getting Started

Prerequisites

• Python 3.11+
• Node.js 18+
• PostgreSQL 15
• Apify account (for scraping)
• Groq API key (for LLM)

1. Clone the repository

git clone https://github.com/your-org/thiqa.git
cd thiqa

2. Backend Setup

cd backend

# Create and activate virtual environment
python -m venv venv
.\venv\Scripts\activate      # Windows
source venv/bin/activate     # Linux/Mac

# Install dependencies
pip install -r requirements.txt

# Configure environment
cp .env.example .env
# Edit .env with your credentials (see Environment Variables below)

# Run database migrations
alembic upgrade head

# Start the backend
uvicorn app.main:app --reload

3. Frontend Setup

cd frontend

# Install dependencies
npm install

# Start the frontend
npm run dev

The app will be available at http://localhost:5173

---

⚙️ Environment Variables

Create a backend/.env file with the following:

DATABASE_URL=postgresql://thiqa_user:your_password@localhost:5432/thiqa_db
SECRET_KEY=your_jwt_secret_key

APIFY_API_KEY=your_apify_token
GROQ_API_KEY=your_groq_api_key
SIGHTENGINE_API_USER=your_sightengine_user
SIGHTENGINE_API_SECRET=your_sightengine_secret
CLOUDINARY_URL=cloudinary://your_cloudinary_credentials

---

📡 API Endpoints

Method	Path	Description	Auth
POST	/api/auth/register/	User registration	No
POST	/api/auth/login/	JWT token issuance	No
GET	/api/search/?q=	Main seller lookup + trust score	No
POST	/api/analyze/text/	Scam text analysis	No
POST	/api/analyze/screenshot/	OCR + scam analysis on images	No
POST	/api/reports/	Submit scam report	Yes
GET	/api/reports/?seller_url=	List reports for a seller	No
POST	/api/reviews/	Submit buyer review	Yes
GET	/api/reviews/?seller_id=	List reviews for a seller	No
GET	/api/blacklist/	Paginated blacklist	No
GET	/api/history/	User search history	Yes

Full interactive docs available at http://localhost:8000/docs

---

📊 Model Performance

MARBERTv2 Sentiment Classifier

• Test Accuracy: 99.2% (375-sample held-out set)
• F1 Macro: 0.992
• Training data: 2,125 samples of Algerian e-commerce text (Arabic/Darija/French)

Gradient Boosting Trust Scorer

• CV F1-Weighted (5-fold): ~0.87
• Training samples: 1,338 (1,296 synthetic + 42 real)
• Top features: weighted_report_score, avg_credibility_score, avg_stars, account_age_days

Scam Text Analyzer (LLM Few-Shot)

• Label Accuracy: 100% (14/14 examples)
• Scam Type Accuracy: 100% (9/9 applicable)

---

👥 Team

Role	Name
AI + Backend Engineer	Amal MESSAOUD
AI + Backend Engineer	Amina MEDABIS
Design + Frontend Engineer	Zhour SACI

---

🏫 About

Built for the I-ACE 2026 Competition by the Intellect Scientific Club at M'Hamed Bougara University of Boumerdès (UMBB), April 2026.

---

📄 License

This project was developed for academic competition purposes. See individual dependency licenses in the technical report.