"The first fraud detection system in the world that speaks Kannada."

Protecting ₹4,000 Crore of small business dreams. One alert at a time.

🚀 Quick Start  •  🏗️ Architecture  •  🧠 ML Engine  •  📂 Project Structure  •  🧪 Testing  •  🎥 Demo Video

---

🛡️ Why We Built This

India is witnessing the world's most successful digital payment revolution via UPI. However, this success has a dark side. Small merchants — the backbone of Bharat — are being targeted by sophisticated fraud syndicates.

Important

The Problem:

• Language Barrier: 100% of existing fraud tools are in English. Rural merchants are left in the dark.
• Latency Gap: Bank-side fraud detection takes minutes. Fraudsters vanish in seconds.
• Cost Barrier: Enterprise security costs thousands of dollars. Small merchants have zero budget.

PaySentinel was born to bridge this gap. We built a system that doesn't just block transactions — it explains them to the merchant in their native tongue, for free, in under 100 milliseconds.

---

🥇 Competitive Analysis

"Why this over existing fraud tools?" This section answers that directly.

Capability	Traditional Bank Fraud Stack	Generic SME Fraud SaaS	PaySentinel
Primary user focus	Large institutions	Online businesses	Indian small merchants (UPI-first)
Language accessibility	Mostly English	Mostly English	Kannada + Hindi + Tamil + Telugu + English
Explainability for non-technical users	Low	Medium	High (plain-language explanation + SHAP reasons)
Voice-based actionability	Rare	Rare	Built-in multilingual voice alerts
Offline fallback for alerts	Rare	Rare	Available (pyttsx3 fallback)
Forensic evidence output	Internal systems only	Limited exports	PDF fraud certificates + QR proof
Real-time simulation + stream pipeline	Internal only	Often add-on	Native Kafka + SSE support
Cost profile for small shops	High	Medium	Open-source and low-cost deployable

Strategic Differentiators

• Bharat-first design: Built for the exact UPI merchant segment that is underserved by English-only fraud tools.
• Trust + action, not just score: Detection, explanation, voice warning, and forensic report in one flow.
• Deployability: Works as local Flask app, containerized production stack, and stream-enabled architecture.
• Enterprise controls: API key protection, rate limiting, signed model cache, security headers, and rotating logs.

---

🏗️ System Architecture

PaySentinel is built as a high-performance, forensic-grade application with a clear separation of concerns.

graph TD
    subgraph "External Sources"
        A[UPI CSV Logs] --> D[Forensic Gateway]
        B[Kafka Live Stream] --> D
    end

    subgraph "PaySentinel Core (Python/Flask)"
        D --> E[Feature Engineering Engine]
        E --> F{Triple ML Ensemble}
        F --> G[Isolation Forest]
        F --> H[One-Class SVM]
        F --> I[Local Outlier Factor]
        
        F --> J[Inference Merger]
        J --> K[SHAP Explainer]
    end

    subgraph "Output Layer"
        K --> L[Regional Voice Engine]
        K --> M[Forensic PDF Generator]
        K --> N[Stark-Tech Dashboard]
        
        L --> O[Kannada/Hindi/Tamil Alerts]
        M --> P[QR-Signed Certificates]
    end

    style D fill:#0d1117,stroke:#00f5d4,stroke-width:2px
    style F fill:#0d1117,stroke:#ff7043,stroke-width:2px
    style K fill:#0d1117,stroke:#ffc107,stroke-width:2px

Loading

---

📂 Project Structure

We follow a clean, modular Frontend-Backend Separation designed for high-speed delivery and auditability.

Pay_Sentinel/
├── src/                           # 🔧 Core Engine
│   ├── app.py                     #    Flask API Gateway & Controller
│   ├── model.py                   #    Triple ML Ensemble Core
│   ├── generate_data.py           #    Forensic Synthetic Data Generator
│   ├── pdf_report.py              #    Bilingual QR-Signed Certificates
│   ├── voice_alerts.py            #    Regional NLP Voice Engine
│   │
│   └── frontend/                  # 🎨 Modern UI Layer
│       ├── templates/             #    Index & Dashboard (Glassmorphism)
│       └── static/                #    Frontend Assets
│           ├── images/            #    HUD, Backgrounds, Section Visuals
│           └── css/               #    Global Styles & Animations
│
├── tests/                         # 🧪 Quality Assurance (99 Tests)
├── models/                        # 💾 Serialized ML Models
├── docs/                          # 📖 Enterprise Documentation
├── Dockerfile                     # 🐳 Production Containerization
└── requirements.txt               # 📦 Pinned Dependencies

---

🛠️ Tech Stack

Component	Technology	Role
Backend		API Orchestration & Security
ML Engine		Triple Ensemble Anomaly Detection
Forensics		AI Interpretability & Vision
Frontend		Glassmorphism & WebGL HUD
Real-time		High-throughput Event Streaming
Linguistic		Regional Multi-language Alerting

---

📊 Data Flow Pipeline

The journey of a transaction from a raw log to a regional voice alert follows a strict, high-speed pipeline.

1. Ingestion & Sanitization

• Raw Data: CSV upload or Kafka stream.
• Validation: Deep check for SQLi, XSS, and mathematical anomalies (negative amounts, extreme outliers).
• Sanitization: Bleach-based HTML stripping for merchant names.

2. Feature Engineering (~15ms)

We transform raw timestamps and amounts into 11 forensic features including velocity engines and merchant fingerprinting.

3. Neural Inference (~30ms)

The heart of PaySentinel is a weighted ensemble of three unsupervised models:

• Isolation Forest (35%)
• One-Class SVM (35%)
• LOF (20%)
• Heuristics (10%)

4. Forensic Output (~50ms)

• Explainability: SHAP identifies the top 4 reasons for the flag.
• Vocalize: Regional voice alerts in Kannada, Hindi, or Tamil.

---

🧠 How It Works

Stage	Process	Technology	Latency
Ingest	Stream / Upload	Flask + Kafka	5ms
Engine	11-Feature Vector	Pandas + NumPy	15ms
Neural	Triple Ensemble	Scikit-Learn	30ms
Explain	SHAP Values	SHAP Library	35ms
Alert	Regional Voice	gTTS / pyttsx3	<100ms

---

⚡ Quick Start

🚀 Local Development

# Clone the vault
git clone https://github.com/Yashaswini-V21/Pay_Sentinel.git
cd Pay_Sentinel

# Install dependencies
pip install -r requirements.txt

# Ignite the engine
python src/app.py

Tip

Visit http://localhost:5000/dashboard to experience the Stark-Tech command center.

---

🧪 Testing & Audit

We maintain a 99% test pass rate across three specialized suites.

pytest tests/ -v --cov=.

---

👨‍💻 Solo Developer

PaySentinel is a passion project built single-handedly for the BLUEPRINT 2026 Hackathon. It represents a journey of learning Forensic ML, Regional NLP, and High-Performance Flask.

#SoloHacker #BharatTech #OpenSource #FintechSecurity #MLOps

"Protecting the backbone of Bharat's economy."

Made with ❤️ by a Solo Learner