Payment-Risk-Exception-AI-Agent

A local application example that demonstrates a hybrid payment-decisioning workflow. A rule based deterministic decision policy first assesess transactions and automatically approves or denies. The transactions that have mismatched info but not severe enough to trigger a denial is routed to an LLM agent which reviews the transactions then approves/denies or sends for human-review.

This example uses the following:

• Stripe's official Python SDK (stripe-python)

• Stripe's official mock server (stripe-mock)

• Rule based deterministic decision policy

• Simulated Merchant Data

• LLM Agent

• MCP Server with discoverable tools

• FastAPI REST endpoints

• SQLite

• Docker

• Postman API testing

• Analyst dashboard

The four demo transactions

1. Everything matches

2. Travel mismatch

3. Compounding risk

4. Mixed evidence

Related MCP server: Payments Agent Gateway (MCP)

LLM Agent MCP flow

For exception transactions, the model itself selects the tools:

LLM
  └─ calls get_transaction_context through MCP
       └─ receives Stripe + merchant context and deterministic signals
  └─ then calls exactly one MCP terminal tool:
       ├─ record_ai_decision (approve or deny)
       └─ create_human_review

The Python orchestrator executes the tool calls requested by the model. It does not silently replace the model with a deterministic fallback and does not choose the terminal MCP action on the model's behalf.

Setup

For testing, please create an .env file and provide the LLM key.

Run:

docker compose up --build

Open:

• Dashboard: http://localhost:8000

• Health/API configuration: http://localhost:8000/api/health

• FastAPI docs: http://localhost:8000/docs

• MCP endpoint: http://localhost:8001/mcp

• Stripe Mock: http://localhost:12111

Important boundaries

• Stripe Mock returns fixture-based Stripe objects. It does not process money, run Stripe risk models, or preserve a real payment lifecycle.

• Merchant session and account data are simulated demo data.

• Clean transactions stay on deterministic policy; the model is called only for exceptions.

• The MCP server allows context retrieval, AI decision persistence, and human-review creation.

• Human-review resolution is intentionally absent from MCP. Only the analyst REST/dashboard path can resolve it.

Screenshots:

Main Screen:

Rules Fired:

AI Explanation:

Transaction Details:

Video Preview:

https://github.com/user-attachments/assets/84ff0226-a9cb-4c60-a4f7-ed301f0b5f5d