A traceable financial modelling engine, built on AWS

A Healthcare & Life Sciences business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had four focus areas, each of which had to be solved for the engine to be trusted by a pilot fund.

Cloud Combinator designed and built the backend engine as four separated concerns: how raw documents become atoms, how atoms are mapped to a canonical chart of accounts, how the numbers are proven internally consistent, and how the model is calculated and changed. The phased plan below keeps each concern independently testable.

Acceptance mirrors the success criteria one for one: extraction and lineage, mapping and review, reconciliation, modelling, the frozen API contract, the architecture, and handover each carry an explicit sign-off condition. The build is scoped at 10 to 12 weeks (projected) and delivers a system engineered for regulated-fund data sensitivity from day one, with encryption via KMS, VPC isolation and per-workspace tenant isolation in place.

• Extraction and lineage — Real documents convert to atom rows with full source lineage (document, page, cell range, snippet, bounding box) and a confidence score per line item.
• Mapping and review — Extracted atoms map to the canonical chart with three-tier routing, auto-map at 0.90 and above, flag for review between 0.70 and 0.89, hold below, with a human-override queue usable end to end.
• Reconciliation — Five default integrity rules run per period, alongside any plain-language custom rule added through the API, and results are reported whether they pass or fail without blocking modelling.
• Modelling — The SaaS three-statement template runs end to end on extracted data, with circular dependencies converging within 50 iterations or being flagged for human review rather than published silently.
• Handover — Monitoring dashboards, runbooks and handover sessions are delivered, with the first pilot fund live on the engine as the closing milestone.

The engine is built to grow beyond its first template without re-architecture. A Google Sheets and Excel live-sync bridge and a template scale-out track are scoped as parallel add-ons, while QuickBooks, Xero and Plaid integrations, SSO, and agent-generated custom models are reserved in the schema for later releases. The result is not just a pilot-ready modelling engine but a foundation the client can extend as it moves from its first fund to many.