PayPal Sentinel | AI-Powered Issue & Fraud Analytics

Workflow
Architecture
The supervisor can route question to one or more agents
Dashboard
WordCloud
Trend analysis
Payment capture
data comparison

Inspiration

Digital payment platforms like PayPal handle millions of transactions and interactions, generating valuable customer feedback. However, reviews often contain scattered information about failed payments, account access problems, refunds, disputes, suspicious activities, and technical issues.

We realized that these reviews are more than just customer opinions — they can reveal recurring problems, potential fraud signals, product weaknesses, and areas requiring immediate attention.

This inspired us to build PayPal Sentinel, an AI-powered issue intelligence and investigation platform that transforms unstructured PayPal app reviews into actionable insights. Our goal is to make customer feedback more meaningful by combining advanced analytics with specialized AI agents

What it does

PayPal Sentinel combines review intelligence, advanced analytics, multi-agent AI investigation, and PayPal Sandbox integration.

1. Review Intelligence & Classification

Performs multi-dimensional classification of PayPal app reviews.
Identifies business context, stakeholders, business operations, and customer intent.
Analyzes product modules, features, workflow stages, and complaint categories.
Extracts reported symptoms, sentiment, and issue patterns.
Provides primary and secondary classification insights with confidence information.

2. Advanced Analytics Dashboard

Our interactive Streamlit dashboard provides:

Review intelligence overview and classification distributions.
Cross-dimensional analysis of customer issues and product features.
Sentiment and complaint analysis.
Temporal trends to understand changing customer concerns.
Text analytics and word-cloud visualizations.

3. Multi-Agent AI Investigation System

We implemented a supervisor-based architecture using four specialized agents:

Complaint Agent: Investigates customer complaints, reported problems, and their potential business impact.
Fraud Agent: Analyzes potential fraud signals and suspicious patterns appearing in customer feedback.
Engineering Agent: Examines technical symptoms, application performance issues, and potential engineering concerns.
Supervisor Agent: Coordinates the specialist agents, routes investigation requests, and consolidates their findings into a unified response.

This architecture enables specialized investigation instead of relying on a single general-purpose chatbot.

4. FastAPI Backend & SQLite Database

REST API for submitting and retrieving customer reviews.
Persistent SQLite storage with upsert support.
Individual review retrieval and health-check endpoints.
Interactive API documentation through Swagger UI.
Foundation for integrating external review sources and data pipelines.

5. PayPal Sandbox Integration

We implemented and successfully tested a complete Sandbox checkout workflow:

OAuth 2.0 authentication.
Payment order creation.
Buyer approval through PayPal-hosted checkout.
Automatic payment capture after approval.
Manual payment capture for testing and fallback.

How we built it

We developed PayPal Sentinel using a modular architecture that separates data ingestion, analytics, AI agents, database operations, and the dashboard.

Technology Stack:

Programming: Python Dashboard: Streamlit AI Frameworks: LangChain, LangGraph LLM Provider: Groq Backend API: FastAPI, Uvicorn Database: SQLite Data Processing: Pandas, NumPy Visualization: Plotly, Matplotlib, WordCloud Machine Learning & Utilities: Scikit-learn, SciPy, BeautifulSoup Payment Integration: PayPal Sandbox REST APIs

LangGraph and LangChain support our multi-agent orchestration, while Groq provides the LLM capabilities. FastAPI handles review ingestion and retrieval, SQLite stores review records, and Streamlit presents the analytics through an interactive dashboard.

Challenges we ran into

Understanding Unstructured Customer Feedback: Customer reviews contain different problems expressed in different ways. Converting them into meaningful classification dimensions was a major challenge.
Multi-Dimensional Classification: Connecting customer complaints with business context, product features, workflow stages, sentiment, and potential business impact required careful structuring of the data.
Multi-Agent Coordination: Designing a supervisor-based architecture that routes requests to the right specialist agent and combines their findings into a coherent response required thoughtful workflow design.
Large-Scale Data Management: Handling large review datasets, expanded classification dimensions, database ingestion, and efficient retrieval required a structured data pipeline.
PayPal API Integration: Implementing OAuth authentication, creating Sandbox orders, handling buyer approval, and capturing payments involved understanding and integrating the PayPal REST API workflow.

Accomplishments that we're proud of

Built a multi-dimensional review intelligence and classification platform.
Developed an interactive dashboard with cross-dimensional, sentiment, temporal, and text analytics.
Implemented a supervisor-based multi-agent AI system with Complaint, Fraud, and Engineering specialist agents.
Integrated LangGraph, LangChain, and Groq for AI-powered investigation workflows.
Developed a FastAPI backend with SQLite-based persistent review storage.
Successfully tested PayPal Sandbox checkout, including OAuth authentication, order creation, buyer approval, and automatic payment capture.
Organized the project into modular components and published the source code, documentation, and MIT License on GitHub.

What we learned

Through this project, we gained practical experience in:

Building multi-agent AI applications using LangGraph and LangChain.
Integrating LLMs through Groq.
Designing supervisor-based agent orchestration and specialized investigation workflows.
Transforming unstructured customer feedback into structured intelligence.
Developing interactive analytics dashboards using Streamlit and Plotly.
Building REST APIs using FastAPI and managing persistent data with SQLite.
Integrating third-party payment services using PayPal Sandbox APIs.
Designing modular AI systems that can be extended with additional capabilities.

What's next for PayPal Sentinel | AI-Powered Issue & Fraud Analytics

We envision evolving PayPal Sentinel into a comprehensive AI-powered support and intelligence ecosystem for both consumers and merchants.

Our future roadmap includes:

Dedicated Consumer AI Assistant: Helping consumers understand payment failures, refunds, disputes, account-related issues, and other PayPal concerns.
Dedicated Merchant AI Assistant: Supporting merchants with payment-related queries, transaction concerns, disputes, and business operations.
Advanced Fraud Intelligence: Improving suspicious-pattern analysis, investigation workflows, and evidence traceability.
Automated Review Intelligence: Triggering classification and relevant AI investigation workflows when new reviews are ingested.
Real-Time Data Integration: Connecting external review sources and feedback pipelines.
Intelligent Escalation: Routing complex or unresolved issues to appropriate support teams.
Security & Scalability: Adding authentication, access control, and deployment infrastructure.

Our long-term vision is to bridge the gap between consumers, merchants, and payment support through intelligent AI-driven assistance.

PayPal Sentinel — Turning Customer Feedback Into Actionable Intelligence.