🛡️ RazorGuard AI — Intelligent Payment Risk Manager
A defense-focused AI payment risk management prototype that analyzes transaction signals, estimates risk, explains detected risk factors, and recommends an appropriate defensive action.
Built for the Razorpay AI Buildathon 2026 — Track 2: AI Risk Manager.
🚀 Overview
RazorGuard AI is an intelligent payment risk management prototype designed to identify potentially risky transactions using behavioral and contextual transaction signals.
The system combines a Random Forest machine learning model with an explainable rule layer to:
Analyze transaction risk
Generate a risk probability score
Classify transactions as LOW, MEDIUM, or HIGH risk
Explain detected risk signals
Estimate the amount at risk
Recommend a defensive action
Maintain a transaction audit trail
Provide model performance and risk analytics
Prototype note: This project uses a synthetic transaction dataset and is intended for demonstration and experimentation, not production fraud detection.
✨ Key Features
🤖 AI Risk Detection
Random Forest-based transaction risk prediction.
📊 Risk Scoring
Each transaction receives a risk score, probability, risk level, and detected risk signals.
🔍 Explainable Risk Signals
The application surfaces signals such as:
Previous chargebacks
Location mismatch
Transaction amount
Account age
Device changes
Transaction frequency
Failed transactions
International activity
🛡️ Automated Defensive Actions
Risk Level
Recommended Action
🟢 LOW
Allow transaction and continue monitoring
🟡 MEDIUM
Step-up verification
🔴 HIGH
Hold for manual review and additional verification
📈 Risk Analytics
Risk distribution
Risk activity trend
Model performance
Confusion matrix
False-positive rate
Top risk signals
False-positive cost simulation
Transaction audit trail
🏗️ System Architecture
Transaction Input │ ▼ Flask REST API │ ▼ Risk Engine │ ├── Feature Processing ├── ML Prediction ├── Risk Scoring └── Risk Explanation │ ▼ Random Forest Model │ ▼ LOW / MEDIUM / HIGH │ ▼ Allow / Verify / Hold
🧠 Machine Learning Model
RazorGuard AI uses a Random Forest Classifier trained on a synthetic dataset containing 10,000 transaction records.
Model Configuration
Algorithm: Random Forest Classifier Estimators: 400 Maximum Depth: 10 Minimum Samples Split: 8 Minimum Samples Leaf: 3 Class Weight: Balanced Max Features: sqrt Random State: 42
Features
amount account_age_days transactions_24h failed_transactions_24h device_changes_30d location_mismatch international_transaction previous_chargebacks
Target:
is_risky
📊 Model Evaluation
The improved model uses a 60% training / 20% validation / 20% untouched test split.
The validation set was used for decision-threshold tuning, while final metrics were measured on the untouched test set.
Final Held-Out Test Performance
Metric
Score
Accuracy
78.80%
Precision
36.65%
Recall
63.45%
F1 Score
46.46%
False Positive Rate
18.60%
Decision Threshold
0.52
Confusion Matrix
Predicted
Not Risky Risky
Actual Not Risky 1392 318 Risky 106 184
Metrics are based on the project's synthetic dataset and held-out test split. They should not be interpreted as real-world or production fraud-detection performance.
🔎 Top Risk Signals
Feature
Importance
Previous Chargebacks
0.3382
Location Mismatch
0.1633
Transaction Amount
0.1269
Account Age
0.1115
Device Changes
0.0963
Transactions in 24h
0.0631
Failed Transactions
0.0614
International Transaction
0.0392
🛡️ Risk Decision Flow
Transaction │ ▼ Feature Analysis │ ▼ ML Risk Probability │ ▼ Risk Score │ ├── < 40 ───────► LOW │ │ │ ▼ │ ALLOW │ ├── 40–69 ──────► MEDIUM │ │ │ ▼ │ STEP-UP VERIFICATION │ └── ≥ 70 ───────► HIGH │ ▼ MANUAL REVIEW
🖥️ Product Screenshots
Dashboard
🔴 High-Risk Detection
🟡 Medium-Risk Action
🟢 Low-Risk Action
📊 Risk Insights & Analytics
🧾 Transaction Audit Trail
🔌 API Endpoints
The backend is implemented using Flask.
Method
Endpoint
Purpose
GET
/
Application landing / health response
GET
/api/health
Backend health check
GET
/api/metrics
Model performance metrics
POST
/api/analyze
Analyze a transaction
GET
/api/audit
Retrieve audit records
DELETE
/api/audit
Clear audit records
Example Request
{ "amount": 1250, "account_age_days": 12, "transactions_24h": 15, "failed_transactions_24h": 4, "device_changes_30d": 3, "location_mismatch": 1, "international_transaction": 1, "previous_chargebacks": 1 }
Example Result
Risk Level: HIGH Risk Score: ~88 Action: HOLD
🧪 Example Risk Scenarios
🔴 HIGH Risk
Multiple elevated-risk signals can result in a HIGH risk classification.
HOLD — Manual review and additional verification required
🟡 MEDIUM Risk
Moderate risk signals can trigger:
STEP-UP VERIFICATION
🟢 LOW Risk
Lower-risk transactions can continue with:
ALLOW — Continue transaction and monitor
📁 Project Structure
RazorGuard-AI/ │ ├── app/ │ ├── app.py │ ├── static/ │ │ ├── script.js │ │ └── style.css │ └── templates/ │ └── index.html │ ├── assets/ │ └── screenshots/ │ ├── dashboard.png │ ├── high-risk-detection.png │ ├── medium-risk-action.png │ ├── low-risk-action.png │ ├── risk-insights-analytics.png │ └── transaction-audit-trail.png │ ├── data/ │ ├── generate_dataset.py │ └── transactions.csv │ ├── models/ │ ├── razorguard_model.joblib │ └── razorguard_improved_model.joblib │ ├── src/ │ ├── train_model.py │ ├── improve_model.py │ ├── evaluate_model.py │ └── risk_engine.py │ ├── .gitignore ├── requirements.txt └── README.md
⚙️ Tech Stack
Backend: Python, Flask, Flask-CORS
Machine Learning: Scikit-learn, Random Forest, NumPy, Pandas, Joblib
Frontend: HTML5, CSS3, JavaScript
Development: Git, GitHub, Python Virtual Environment
🚀 Getting Started
- Clone the Repository
git clone https://github.com/Ashutosh9-pan/RazorGuard-AI.git cd RazorGuard-AI
- Create a Virtual Environment
Windows
python -m venv venv .\venv\Scripts\activate
- Install Dependencies
pip install -r requirements.txt
- Run the Application
python app\app.py
Open:
🧪 Model Training
Generate the synthetic dataset:
python data\generate_dataset.py
Train the baseline model:
python src\train_model.py
Train the improved model:
python src\improve_model.py
Evaluate the model:
python src\evaluate_model.py
💰 False-Positive Awareness
Payment risk systems need to balance security with customer experience.
A model that flags too many legitimate transactions can create unnecessary friction, while a model that misses risky transactions can increase exposure.
RazorGuard AI therefore exposes:
Precision
Recall
F1 Score
False Positive Rate
Confusion Matrix
Decision Threshold
False-Positive Cost Simulation
This project is a prototype:
The dataset is synthetic.
The model has not been validated on real payment data.
Real-world risk patterns can change over time.
Risk rules are prototype heuristics.
The audit trail is currently maintained in application memory.
The model should not be considered production-ready fraud detection.
A production implementation would require additional validation, monitoring, privacy controls, security hardening, model governance, and evaluation on representative real-world data.
🔮 Future Improvements
Real-time event streaming
Persistent transaction storage
Model monitoring and drift detection
Automated threshold optimization
Advanced anomaly detection
Ensemble risk models
Explainable AI techniques
Authentication and role-based access
Production-grade observability
Human-review feedback loops
Integration with payment processing systems
🎯 Buildathon Objective
RazorGuard AI was developed for:
Razorpay AI Buildathon 2026 — Track 2: AI Risk Manager
The project demonstrates a working AI-assisted risk management workflow:
Detect → Explain → Score → Decide → Respond
with measurable model performance and transparent limitations.
👨💻 Author
Ashutosh Panwar
B.Tech CSE Graduate AI/ML Developer | Data Analyst | Android Developer
GitHub: Ashutosh9-pan
Portfolio: ashutosh-panwar-portfolio.vercel.app
⭐ Project Summary
RazorGuard AI demonstrates how machine learning, explainable risk signals, and automated defensive actions can be combined into a single payment risk management workflow.
Transaction ↓ AI Risk Analysis ↓ Explainable Signals ↓ Risk Classification ↓ Defensive Action ↓ Audit & Analytics
Built as a prototype for the Razorpay AI Buildathon 2026.