🔍 SatyaSetu.AI

AI-Powered Digital Forensics & Fraud Detection Platform

SatyaSetu.AI is an intelligent forensic analysis platform that combines OCR, NLP, OSINT, and Machine Learning to detect fraud patterns in digital evidence and public procurement contracts.

---

🏗️ Architecture Overview

┌─────────────────────────────────────────────────────────────────────────────┐
│                              SatyaSetu.AI                                      │
├──────────────────────────────┬──────────────────────────────────────────────┤
│        FRONTEND              │               BACKEND (FastAPI)              │
│        (Next.js)             │                                              │
│                              │  ┌─────────────────────────────────────────┐ │
│  • Evidence Upload           │  │           AI PIPELINES                  │ │
│  • Case Dashboard            │  │                                         │ │
│  • Entity Intelligence       │  │  📄 OCR Engine (Tesseract + PyMuPDF)    │ │
│  • Fraud Prediction UI       │  │  🧠 NER (spaCy en_core_web_sm)          │ │
│  • Batch Analysis            │  │  🎯 Scam Classifier (TF-IDF + SBERT)    │ │
│  • Report Generation         │  │  🔗 URL/QR Scanner + OSINT              │ │
│                              │  │  ⚖️ Fraud Predictor (XGBoost)           │ │
│                              │  │  📊 Risk Assessor                       │ │
│                              │  └─────────────────────────────────────────┘ │
│                              │                                              │
│                              │  ┌─────────────────────────────────────────┐ │
│                              │  │           ML MODELS                     │ │
│                              │  │                                         │ │
│                              │  │  • fraud_detection_model.pkl (XGBoost)  │ │
│                              │  │  • scam_classifier.pkl (Logistic Reg)   │ │
│                              │  │  • tfidf_vectorizer.pkl                 │ │
│                              │  │  • target_encoders.pkl                  │ │
│                              │  └─────────────────────────────────────────┘ │
└──────────────────────────────┴──────────────────────────────────────────────┘

---

🚀 Features

📄 Evidence Processing Pipeline

Step	Module	Description
1	Upload Evidence	Secure file upload with chain of custody logging
2	OCR Engine	Extract text from images/PDFs using Tesseract & PyMuPDF
3	NER Extraction	Identify entities (names, orgs, locations) using spaCy
4	Scam Classification	AI-powered scam detection with confidence scoring
5	URL/QR Scanning	Extract and analyze URLs with OSINT enrichment
6	Risk Assessment	Comprehensive threat scoring and recommendations
7	Report Generation	PDF forensic reports with all findings

⚖️ Fraud Detection (Public Procurement)

AI-powered corruption risk prediction for government contracts.

---

🤖 Machine Learning Models

1. 🚨 Fraud Detection Model (XGBoost)

Purpose: Predict corruption risk in public procurement contracts

Training Data:

• Source: Digiwhist Romania 2023 (Open Contracting Data)
• Size: 1.8M+ public procurement contracts
• Target Variable: Composite Risk Indicator (CRI) — a score from 0-1

Model Performance:

Metric	Value
R² Score	0.74
Test RMSE	0.098
Train RMSE	0.091

Key Features Used:

Feature	Description
tender_finalprice	Final awarded contract value
tender_estimatedprice	Initial estimated value
tender_recordedbidscount	Number of bidders
price_efficiency	Ratio: final_price / estimated_price
is_round_1000	Flag if final price is divisible by 1000
single_bidder_proxy	Flag if only 1 bidder participated
title_length	Length of contract title
is_medium_title	Flag if title is 100-200 characters
is_sunday	Flag if awarded on Sunday
is_december	Flag if awarded in December
buyer_encoded	Target-encoded buyer organization
winner_encoded	Target-encoded winning company

Rule-Based Fraud Signals:

🚩 Round Number Trap    — Final price divisible by 1000
🚩 Single Bidder        — Only one bidder (bid rigging indicator)
🚩 Vague Title          — Contract title too short (<30 chars)
🚩 Cost Overrun         — Final price exceeds estimate by >5%
🚩 Sunday Award         — Unusual timing (awarded on Sunday)
🚩 December Rush        — Year-end budget spending rush

Risk Levels:

CRI Score	Level	Action
≥ 0.70	🔴 CRITICAL	Immediate investigation required
≥ 0.50	🟠 HIGH	Detailed audit recommended
≥ 0.30	🟡 MODERATE	Enhanced monitoring advised
< 0.30	🟢 LOW	Standard oversight sufficient

---

2. 🎯 Scam Classifier

Purpose: Classify text evidence as potential scam/fraud

Architecture: Hybrid approach combining:

• TF-IDF Vectorizer — Text feature extraction
• Logistic Regression — Primary classifier
• Sentence-BERT — Semantic similarity matching to known scam patterns

Scam Categories Detected:

• Phishing / Fake Bank Emails
• KYC Verification Scams
• Lottery/Prize Scams
• Investment Fraud
• Tech Support Scams
• Romance Scams

---

📂 Project Structure

SatyaSeth.AI/
├── backend/                    # FastAPI Backend
│   ├── app/
│   │   ├── api/               # API Routes
│   │   │   ├── upload_evidence.py
│   │   │   ├── analyze.py
│   │   │   ├── fraud_predict.py   # 🚨 Fraud Detection API
│   │   │   ├── report.py
│   │   │   ├── batch_analyze.py
│   │   │   └── threat_hub.py
│   │   ├── pipelines/         # AI Processing Pipelines
│   │   │   ├── ocr.py             # Tesseract + PyMuPDF
│   │   │   ├── ner.py             # spaCy NER
│   │   │   ├── scam_classifier.py # ML Scam Detection
│   │   │   ├── url_qr_scanner.py  # URL/QR Extraction
│   │   │   ├── osint_engine.py    # OSINT Lookups
│   │   │   ├── risk_assessor.py   # Risk Scoring
│   │   │   └── report_generator.py
│   │   ├── models/            # Trained ML Models
│   │   │   ├── fraud_detection_model.pkl  # XGBoost (Romania)
│   │   │   ├── scam_classifier.pkl
│   │   │   ├── tfidf_vectorizer.pkl
│   │   │   └── target_encoders.pkl
│   │   └── main.py            # FastAPI App Entry
│   ├── Dockerfile
│   └── requirements.txt
│
├── webapp/                    # Next.js Frontend
│   ├── src/
│   │   ├── app/
│   │   │   ├── fraud-predict/    # Fraud Prediction UI
│   │   │   ├── entities/         # Entity Intelligence
│   │   │   └── ...
│   │   └── lib/
│   │       └── api.ts            # API Client
│   └── package.json
│
└── docker-compose.yml

---

🛠️ Tech Stack

Layer	Technology
Frontend	Next.js 14, React, TailwindCSS, Framer Motion
Backend	FastAPI, Python 3.10, Uvicorn
ML/AI	XGBoost, scikit-learn, spaCy, Sentence-Transformers
OCR	Tesseract, PyMuPDF
Database	File-based (JSON/Pickle for MVP)
Deployment	Docker, Render

---

🚀 Quick Start

Prerequisites

• Python 3.10+
• Node.js 18+
• Docker (optional)

Backend Setup

cd backend

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: .\venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

# Download spaCy model
python -m spacy download en_core_web_sm

# Run server
uvicorn app.main:app --reload

Frontend Setup

cd webapp

# Install dependencies
npm install

# Create .env.local
echo "NEXT_PUBLIC_API_URL=http://127.0.0.1:8000/api" > .env.local

# Run development server
npm run dev

Docker Deployment

# Build and run with Docker Compose
docker-compose up --build

---

📡 API Endpoints

Core Endpoints

Method	Endpoint	Description
POST	/api/upload-evidence	Upload evidence files
POST	/api/analyze	Full forensic analysis
GET	/api/report/{file_id}	Generate PDF report

Fraud Detection Endpoints

Method	Endpoint	Description
POST	/api/fraud-predict	Predict single contract risk
POST	/api/fraud-predict/batch	Batch contract analysis
GET	/api/fraud-predict/model-info	Model metadata

Example: Fraud Prediction Request

curl -X POST "http://localhost:8000/api/fraud-predict" \
  -H "Content-Type: application/json" \
  -d '{
    "name": "Road Construction Phase 1",
    "department": "Ministry of Transport",
    "estimated_price": 500000,
    "final_price": 550000,
    "bidders": 1,
    "award_month": 12,
    "is_sunday": false,
    "is_december": true
  }'

---

📊 Model Training (Fraud Detection)

The fraud detection model was trained on public procurement data:

# Training notebook: notebooks/fraud_detection_training.ipynb

# Data source
data = pd.read_csv("data-romania-2023.csv")  # 1.8M+ contracts

# Target variable
# Composite Risk Indicator (CRI) — combines multiple red flags

# Model: XGBoost Regressor
model = XGBRegressor(
    n_estimators=200,
    max_depth=6,
    learning_rate=0.1,
    subsample=0.8,
    colsample_bytree=0.8
)

# Cross-validated on Belgium data for generalization

---

🔒 Security Notes

• CORS configured for specific origins in production
• File uploads validated and sanitized
• Chain of custody logging for evidence
• No sensitive data stored in version control

---

👥 Team

Escape Da Vinci — Building AI for transparent governance

---

📄 License

MIT License — See LICENSE for details.

---

🙏 Acknowledgments

• Digiwhist / OpenTender — Public procurement data
• spaCy — NLP pipeline
• XGBoost — Gradient boosting framework
• FastAPI — Modern Python web framework