Repository navigation
Show and tell: SentinelPay — fail-safe Spring AI fraud analysis in an event-driven payment pipeline #7224
Replies: 1 comment
|
The three-layer decision boundary is a strong design choice: deterministic controls own hard decisions, while the model is an evidence-producing component for the ambiguous cases. Routing uncertainty to For the three areas you called out, I would consider:
I would also propagate a transaction/correlation id through Kafka, the AI call, and the review event, then retain the exact prompt version and normalized model output. That gives an auditor a reproducible explanation for why a particular transaction was reviewed without making the model responsible for the final payment decision. Great use case for Spring AI as an assistive layer rather than an autonomous authority. |
Uh oh!
There was an error while loading. Please reload this page.
I built SentinelPay, an open-source UPI payment simulation that uses Spring AI as the third layer of a fraud-verdict pipeline.
The design deliberately does not let the model make the first or only decision:
ChatClientevaluates transactions that did not receive a hard fraud verdict.The Spring AI path uses a structured
VERDICT / CONFIDENCE / REASONresponse, a dedicated executor, and a five-second timeout. Empty responses, timeouts, parsing failures, and provider errors fail safely toREVIEWfor human handling—never automatically toSAFE.The broader system uses Spring Boot microservices, Kafka for transaction/fraud events, Redis, MySQL, Docker Compose, Prometheus, Grafana, and a React dashboard.
I would especially value feedback from the Spring AI community on:
Repository: https://github.com/thillainirmal-tech/SentinelPay
Live demo: https://sentinelpay.thillainirmal-tech.dev/#/login
Related paper: https://ijracse.org/index.php/cse/article/view/2477
Feedback and focused contributions are welcome.
All reactions