Governed accounts-payable agents for a bank

The scenario

InvoiceSentinel is a finance-operations platform where AI agents do real work under real controls. Invoices are extracted, matched against purchase orders, scored for risk and routed to the right approver; low-risk ones are paid automatically. Around the agent runtime sit identity, data, model, policy, runtime and operations governance — implemented with open-source components in a single Docker stack, calling real LLMs through a gateway.

System architecture

High-level flow of InvoiceSentinel, from input to output.

What you will learn

• Identity for both humans and agents
• PII detection and data governance in an LLM pipeline
• Policy-as-code for what an agent may do
• Model registry, tracing and operational dashboards

What you will build

• Invoice extraction and PO-matching agents
• Risk scoring and approval routing
• Governance layers as a reproducible Docker stack
• Audit dashboard proving who (or what) did each action

Build plan

1. Invoice extraction and matching
2. Risk scoring and routing
3. Identity, PII and policy layers
4. Observability and audit dashboard
5. End-to-end scenario run

Tech stack