FraudOps Agent

InspirationFraudOps Agent

Inspiration

Financial analysts spend a significant amount of time reviewing alerts, gathering evidence from multiple systems, writing investigation notes, and updating case management platforms. While large language models are becoming increasingly capable, many AI demos stop at answering questions instead of completing real operational workflows.

We wanted to build an agent that behaves like a junior analyst: collecting information, analyzing transaction activity, drafting an investigation memo, and then handing control back to a human reviewer before any action is taken.

Our goal was not to build a fraud detection system. Instead, we focused on building an explainable review workflow that combines deterministic analytics, agent reasoning, and human oversight.

⸻

What It Does

FraudOps Agent is an AI-powered alert review assistant built for synthetic financial investigations.

Given an alert such as A-1027, the agent:

1. Retrieves alert details from MongoDB through MongoDB MCP.
2. Retrieves customer, account, and transaction information.
3. Computes transaction network features such as fan-in, fan-out, pass-through behavior, and repeated transaction patterns.
4. Uses Gemini to generate a structured analyst memo.
5. Produces evidence-backed explanations and recommended next steps.
6. Waits for human approval.
7. Writes the memo, follow-up tasks, and updated case status back to MongoDB.

The system never makes a final fraud determination and always requires human review.

⸻

How We Built It

Frontend

• Next.js
• React

Backend

• FastAPI
• Python

Agent Layer

• Google ADK
• Gemini

Data Layer

• MongoDB Atlas
• MongoDB MCP

Analytics Layer

• NetworkX
• Custom graph feature engine

Deployment

• Google Cloud Run

The backend computes deterministic graph features before passing evidence to Gemini. This ensures the model explains findings instead of inventing them.

⸻

Challenges We Ran Into

One of the biggest challenges was preventing the model from acting as a fraud detector.

We wanted the system to assist analysts rather than replace them. To achieve this, we implemented strict guardrails that prevent the agent from making definitive claims such as “this is fraud” or “this is money laundering.”

Another challenge was separating deterministic analytics from reasoning. Instead of asking Gemini to infer graph metrics directly, we built a graph feature engine that computes network characteristics and passes structured evidence into the model.

Finally, integrating retrieval, reasoning, memo generation, and write-back actions into a single workflow required careful orchestration through the agent layer.

⸻

Accomplishments That We’re Proud Of

• Built a complete end-to-end agent workflow rather than a chatbot.
• Integrated MongoDB MCP as the primary retrieval and action layer.
• Combined graph analytics with Gemini reasoning.
• Implemented human-in-the-loop approval before any system updates.
• Generated evidence-backed analyst memos using synthetic financial data.
• Successfully updated case records and workflow status through the agent.

⸻

What We Learned

This project reinforced the importance of combining deterministic systems with AI reasoning.

We learned that agents become significantly more useful when they can complete operational workflows instead of simply generating text.

We also learned how MCP can serve as a powerful bridge between agents and enterprise systems, enabling both retrieval and action in a controlled manner.

⸻

What’s Next

Future versions of FraudOps Agent could support:

• Additional alert types
• More advanced transaction graph analytics
• Multi-agent investigation workflows
• Analyst feedback learning loops
• Real-time case prioritization
• Enterprise audit and governance features

Our long-term vision is to transform alert review from a manual investigation process into a transparent, AI-assisted workflow with human oversight at every critical decision point.

Note: All data used in this project is synthetic and generated for demonstration purposes only.

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for FraudOps Agent

Built With

• agentic-ai
• fastapi
• generative-ai
• google-adk
• google-cloud-run
• google-gemini
• mongodb
• mongodb-atlas
• mongodb-mcp
• networkx
• next.js
• python
• react
• rest-apis
• synthetic
• typescript