The scenario
Most agent demos are a chatbot with a clever prompt. RefundRadar takes one real operational process — reviewing every return and refund request — and hands it to a system of agents that is bounded, observable and reversible. An intake agent normalises the request, a policy agent checks eligibility, a risk agent scores fraud signals, and a decision agent either clears it or escalates to a reviewer dashboard. Every decision is traceable to the model, prompt version, policy version, inputs, tokens and cost that produced it — the kind of system a Forward Deployed Engineer is expected to hand over.
System architecture
01 Return request Order, reason, photos
02 Intake agent Normalise + enrich
03 Policy agent Eligibility rules
04 Risk agent Fraud signals + score
05 Decision agent Approve or escalate
06 Reviewer desk Human-in-the-loop
High-level flow of RefundRadar, from input to output.
What you will learn
- Splitting a business process into agents with narrow, auditable responsibilities
- Bounded autonomy: spend limits, confidence thresholds and human escalation
- Per-agent identity and decision logging for full traceability
- Scenario-driven validation with synthetic fraud and edge cases
- Writing the business case and the handover document
What you will build
- Four-agent decision pipeline with a policy engine the business can edit
- Reviewer dashboard for escalated cases with the agent’s reasoning trail
- Decision ledger storing model, prompt, policy and cost for every outcome
- Scenario test harness covering 40 return situations
Build plan
- Model the returns process and the policy rules
- Build intake, policy and risk agents with tests
- Add the decision agent with escalation thresholds
- Ship the reviewer dashboard and decision ledger
- Run the scenario suite and write the handover