DETECT • UNDERSTAND • PROTECT
Designed for the Maxim Code AI Software Innovation Competition.
- 🌐 Web Demo Interface: https://fraudshield-ai-q7ug.onrender.com/demo/
- 📄 API Documentation (Swagger UI): https://fraudshield-ai-q7ug.onrender.com/docs
- ⚙️ Production Backend API: https://fraudshield-ai-q7ug.onrender.com
India is experiencing a massive increase in cyber fraud, particularly "Digital Arrest" scams. In these scams, fraudsters impersonate government agencies, police, CBI, tax authorities, or banks, psychologically pressuring victims under the threat of arrest or account freezing.
The core challenge isn't simply a lack of awareness; it is timing. Most victims realize they have been scammed only AFTER their money has been transferred. There is a critical lack of real-time intervention at the exact moment the suspicious communication is taking place.
FraudShield AI is an AI-powered real-time scam detection and user protection platform. It provides an immediate warning at the moment suspicious communication is analyzed.
Instead of acting as a generic educational chatbot or a post-facto security dashboard, FraudShield AI focuses entirely on "Intervention before financial loss."
- Innovation: Real-time scam intelligence pipeline instead of a generic chat interface.
- Problem Solving: Focuses on the exact moment of pressure to intercept financial transactions.
- Explainable AI (XAI): Displays why a message was flagged, listing specific red flags, a risk score, and plain-language reasoning.
- Technical Skills: Combines Android (Kotlin + Compose), Python (FastAPI), Speech-to-Text abstractions, rule-based heuristics, and configurable LLMs.
- SMS & WhatsApp Scan: Copy-paste or upload messages for analysis.
- Call Transcript Scan: Paste or simulate a call conversation for scanning.
- Audio File Upload: Speech-to-Text transcription followed by scam pattern recognition.
- Scam Classification: Classifies text into 8 distinct scam categories.
- Explainable Warning Alerts: Prominent danger screen with detailed red flags.
- Reporting Interface: One-click action to guide reporting the scam to the National Cyber Crime Helpline (1930).
- Scan History: Stores previous scans locally and optionally to Firebase Firestore.
- Interactive Web & Mobile App: Demo web portal and full Jetpack Compose Android client.
The platform implements a distinct four-stage workflow:
- INPUT (Detect): The user submits suspicious communication (text, transcript, or audio).
- ANALYSIS (Understand): The system normalizes the input and applies Speech-to-Text (STT) + Natural Language Processing (NLP).
- SCAM INTELLIGENCE (Classify): Real-time checks extract scam patterns, classify the category, and calculate an explainable risk score.
- PROTECTION (Protect): The user receives a detailed warning dashboard, highlighted red flags, and immediate safety recommendations (e.g. "No Indian agency conducts arrests via video call. Do not transfer funds.").
FraudShield AI operates a hybrid intelligence layer:
Context Analysis ──> Pattern Detection ──> Scam Classification ──> Risk Scoring ──> Explainable Alert
Rather than relying blindly on an LLM-generated risk score, the system calculates score factors deterministically based on key security indicators:
- Government / Law Enforcement Impersonation: +25 points
- Arrest / Legal Threat: +25 points
- Urgent Payment / Money Demand: +20 points
- OTP / PIN / Password Request: +20 points
- Fear / Psychological Coercion: +15 points
- Suspicious URL / Link: +15 points
- Investment / Returns Guarantee: +20 points
- Prepaid Task / Job Fraud: +20 points
- Bank / KYC Impersonation: +20 points
- Lottery / Prize / Reward Scam: +18 points
- Aadhaar / Identity Threat: +18 points
- UPI / Payment Fraud: +18 points
- 0–29: LOW Risk
- 30–59: MEDIUM Risk
- 60–79: HIGH Risk
- 80–100: CRITICAL Risk
- Backend: Python 3.13, FastAPI, Uvicorn, Pydantic, Pytest
- Mobile Application: Android Kotlin, Jetpack Compose, Material 3, Retrofit, OkHttp, Room Database, Jetpack Navigation
- Web Demo Interface: Vanilla HTML5, CSS3, ES6 JavaScript
- Cloud Infrastructure: Firebase Firestore (optional storage)
- AI Integrations: Configurable LLM API (Google Gemini, OpenAI GPT, or Fallback Engine)
- Digital Arrest Scam
- Fake Bank Support
- Investment Scam
- Prepaid Task / Job Scam
- Phishing
- UPI / Payment Scam
- Government / Authority Impersonation
- Lottery / Prize Scam
The project contains 6 preset demo scenarios in the database:
- Digital Arrest Scenario (CRITICAL): "CBI" video call money-laundering threat.
- Fake Bank Support (CRITICAL): "SBI" customer care KYC block threat + OTP request.
- Investment Scam (HIGH): 200% guaranteed return Telegram stock group invitation.
- Prepaid Task Scam (HIGH): Earn ₹500/task with upfront registration fee.
- Phishing SMS (HIGH): Urgent SBI account block notification link.
- Safe Message (LOW): Normal dinner invitation.
FraudShield-AI/
│
├── android-app/ # Kotlin + Compose Android app
│ ├── app/ # Jetpack Compose UI, Retrofit API client
│ └── README.md # Mobile app documentation
│
├── backend/ # FastAPI server codebase
│ ├── app/
│ │ ├── main.py # FastAPI server entry point & static mounts
│ │ ├── api/ # API Router endpoints
│ │ ├── models/ # Pydantic schemas
│ │ ├── services/ # Firebase store, STT service, demo data
│ │ ├── ai/ # Configurable LLM and Rule Fallback Engine
│ │ └── scoring/ # Heuristics & risk scoring engines
│ ├── requirements.txt # Python backend dependencies
│ └── .env.example # Environment template configuration
│
├── demo-web/ # HTML + CSS + JS developer/judging interface
│ ├── index.html
│ ├── style.css
│ └── app.js
│
├── docs/ # Deep-dive architecture and API specifications
│ ├── architecture.md
│ ├── api.md
│ └── demo-scenarios.md
│
├── README.md # Root product overview & instructions
└── LICENSE # MIT License
Configure the backend by creating a .env file in the backend/ folder:
# Copy example
cp backend/.env.example backend/.envFill in the parameters (LLM configurations are optional; the fallback engine will run automatically if left empty):
AI_PROVIDER=fallback
STT_PROVIDER=demo
FIREBASE_ENABLED=false- Navigate to the backend directory:
cd backend
- Set up the virtual environment:
python -m venv venv .\venv\Scripts\activate - Install dependencies:
pip install -r requirements.txt - Start the FastAPI server:
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload
- View API docs at:
http://localhost:8000/docs - Access the Web Demo interface directly at:
http://localhost:8000/demo/
- Open the
/android-appfolder inside Android Studio. - Ensure Gradle synchronization completes successfully.
- Choose a device (or emulator) running Android API 26 or higher.
- Modify
android-app/app/build.gradle.ktsAPI_BASE_URLto point to your machine's local IP address (orhttp://10.0.2.2:8000for default Android emulators). - Compile and run the application.
- Technical Honesty: FraudShield AI acts as an analyzer for user-submitted communications (text, files, transcripts). It does NOT intercept live phone calls or intercept WhatsApp data directly due to Android sandbox and permission models. The architecture is modular to allow integration with call-screen APIs later.
- False Positives: The application displays a prominent warning disclaimer: "AI-generated risk assessment. Verify critical decisions through official channels." It does not make absolute legal claims.
- Privacy First: Sensitive data is processed transiently and is not stored unless a user explicitly saves a scan to their local history.
- Project: FraudShield AI
- Competition: Maxim Code AI Software Innovation Competition
- Team: Students Innovation Team
- License: MIT License