Kumar starts with a declined laptop purchase to explain why a conventional rule or machine-learning score may block a legitimate transaction without giving customer support a clear reason. His proposed design retains existing checks as tier one and sends ambiguous cases to a second tier of agents. It is presented as an addition to established business systems, not a replacement for every fraud decision.
The architecture separates transaction, account, device and payment data into bounded contexts. Domain events and change feeds or message brokers feed read-optimized projections, creating a contextual view for the agent layer. A risk agent and a behavior agent inspect that view through defined tools; a third verdict agent combines their findings and emits an event back into the wider payment workflow.
Kumar notes a sub-500-millisecond transaction requirement and the need to bound an agent's reasoning loop. A closing proof of concept uses synthetic events, but the database starts slowly and the captions do not show a completed end-to-end verdict. The transcript therefore supports the architecture and intended workflow, not measured fraud accuracy, latency or production readiness.
Watch on YouTube
