Turning days of underwriting into hours

A FinTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas, each essential to a system that underwriters would actually trust.

An application arrives at the system through Amazon API Gateway, which passes the proposal to a Supervisor Agent running on Amazon Bedrock AgentCore. The Supervisor opens an AgentCore Memory session for that proposal and orchestrates the specialised agents beneath it. A Database Query Agent gathers the structured application data from the client's Postgres and MongoDB stores, and a Document Agent pulls supporting PDFs from Amazon S3, using vision capability to read scanned bank statements and accounts and to flag anything missing or unreliable.

The engagement delivered a working multi-agent underwriting system on the client's AWS environment, meeting the success criteria set out in the Statement of Work: an active AgentCore Memory resource, all specialised agents operational, end-to-end orchestration, and infrastructure provisioned as code. The following figures describe the design point and the projected operating scale rather than an approved production benchmark.

• Parallel, multi-theme assessment — Business, director, property and affordability agents run together under a Supervisor, so a full application is assessed in one coordinated pass rather than sequential manual review.
• Lender-specific learning, safely isolated — AgentCore Memory captures each lender's acceptance patterns under a dedicated namespace, improving recommendations over time while keeping every tenant's data fully separated.
• Fewer avoidable rejections — The system surfaces valid mitigations from proven language where offsetting factors exist, helping underwriters approve sound applications that a single flag might otherwise have sunk.
• Explainable by design — Every recommendation carries an executive summary, key flags and confidence scores, giving underwriters a defensible basis for the credit paper they sign.
• Handover — The client received the solution as infrastructure as code with technical documentation, architecture diagrams and knowledge-transfer sessions, ready for continued development in their own account.

With the agent chain, memory and infrastructure in place, the client has a foundation that can extend to more assessment themes and additional lenders, taking origination intelligence to tier-one banks that need decisions in hours rather than days.