AgentShield

Dashboard
Transactions
Fraud networks
AI Investigation centre
Payment operations
Audit centre
Investigation workspace

Inspiration

Digital payment fraud is becoming increasingly sophisticated. Fraudsters no longer operate through isolated suspicious transactions; they use networks of accounts, devices, IP addresses, and coordinated behaviors to evade traditional detection systems.

We identified a major gap in existing fraud prevention workflows: detecting a suspicious transaction is only the beginning. Financial institutions also need to understand why it is suspicious, uncover its connections, investigate supporting evidence, and make safe decisions quickly.

We built AgentShield to transform fraud detection from simple transaction flagging into intelligent, explainable, multi-agent fraud investigation.

What it does

AgentShield is an AI-powered, multi-agent fraud detection and investigation platform designed to identify suspicious payment activity and coordinate the investigation process.

Its key capabilities include:

Multi-Agent Intelligence: Specialized agents analyze transaction behavior, identity signals, relationships, and fraud risks.

Fraud Network Explorer: Visualizes connections between suspicious accounts, transactions, devices, payment identities, and IP addresses to uncover coordinated fraud patterns.

Explainable Risk Assessment: Generates risk scores and evidence-backed explanations showing why a transaction may be suspicious.

Investigation Workspace: Organizes investigations, evidence, agent findings, timelines, and case statuses in one centralized interface.

Human-in-the-Loop Authorization: Ensures sensitive financial actions require authorized human approval.

Payment Operations: Supports payment and refund workflows through PayPal integration, with clearly separated simulation and Sandbox modes.

Audit Center: Maintains a chronological record of investigation activities, decisions, approvals, and payment operations.

Our goal is to connect fraud detection, investigation, and financial response into one unified workflow.

How we built it

We built AgentShield using a modular full-stack architecture.

Frontend: Next.js and React for the interactive enterprise dashboard and investigation workspace.

Data Visualization: AG Grid for transaction management and interactive fraud network visualization.

Backend: Python FastAPI for API development, investigation orchestration, and business logic.

Fraud Graph Engine: NetworkX for modeling relationships between transactions, accounts, devices, and other entities.

AI Architecture: A multi-agent orchestration design with specialized investigation stages and deterministic risk assessment.

Payment Integration: PayPal Developer Platform and Sandbox APIs for payment and refund workflows.

Audit System: Structured event tracking for investigation decisions and financial operations.

We designed the architecture to keep deterministic fraud evaluation separate from AI-generated explanations, while maintaining human authorization for sensitive financial operations.

Challenges we ran into

One of our biggest challenges was maintaining consistency across multiple investigation agents and interconnected fraud evidence.

A suspicious transaction can be connected to multiple accounts, devices, IP addresses, and payment identities. Ensuring that all agents reference the same underlying evidence required careful data modeling and relationship management.

Other challenges included:

Preventing overlapping nodes and unreadable relationships in complex fraud graphs.

Maintaining consistent transaction, case, and payment references across different modules.

Designing reliable investigation state transitions.

Preventing duplicate financial operations.

Separating simulated payment operations from actual PayPal Sandbox responses.

Ensuring AI-generated explanations never become fabricated evidence or independently authorize financial actions.

These challenges helped us prioritize reliability, explainability, and financial safety alongside innovation.

Accomplishments that we're proud of

We are proud to have built a unified fraud investigation platform rather than another basic AI chatbot or transaction classifier.

Our key accomplishments include:

Developing a specialized multi-agent fraud investigation architecture.

Building an interactive relationship graph to visualize connected fraud indicators.

Creating an evidence-driven investigation workspace with risk assessments and investigation timelines.

Integrating human approval into sensitive financial workflows.

Building a centralized audit center for traceability and accountability.

Developing a dedicated payment operations module with clear simulation and integration states.

Connecting multiple investigation modules through a unified enterprise dashboard.

Most importantly, AgentShield is designed to explain not only that a transaction is suspicious, but also why it deserves investigation.

What we learned

We learned that effective AI-powered fraud prevention requires more than sophisticated models.

Reliable evidence, consistent data relationships, explainability, deterministic validation, and human oversight are equally important.

We also learned that multi-agent systems need clearly defined responsibilities and shared sources of truth. Without those safeguards, multiple agents can produce inconsistent conclusions.

Finally, we learned that financial automation must be designed around trust: every decision should be traceable, every sensitive action should be authorized, and every payment operation should be independently verified.

What's next for AgentShield

Our vision is to evolve AgentShield into a production-ready intelligent fraud investigation and prevention platform.

Our next steps include:

Integrating and validating live AI providers for evidence analysis and investigation reasoning.

Completing end-to-end PayPal Sandbox payment and refund verification.

Improving fraud graph scalability for larger and more complex networks.

Introducing adaptive fraud detection using historical transaction patterns.

Adding real-time transaction monitoring and suspicious activity alerts.

Improving investigation collaboration and analyst workflows.

Developing advanced fraud pattern discovery and explainable risk models.

Exploring integration with additional payment providers and financial systems.

AgentShield aims to move financial security beyond detecting suspicious transactions toward understanding, investigating, and responding to fraud intelligently and responsibly.