Building an AI-augmented operations platform for an Australian fintech.

The challenge

An Australian fintech and AI software company needed to replace a patchwork of spreadsheets, email folders, and manual processes with a unified operations platform. The system needed to handle client management, document workflows, compliance tracking, and operational reporting.

The interesting constraint: the business handles documents that require careful classification and extraction, and mistakes have consequences. Any AI-assisted processing needed human oversight on consequential decisions.

What we built

The platform has three main layers:

Client and case management

A CRM-like system for managing client relationships and tracking cases through their lifecycle. Each case has a timeline, document attachments, status tracking, and audit history. The system enforces business rules about what can happen when and who needs to approve it.

Document workflow

Documents arrive from multiple channels: email, upload, API integrations. The system routes them to the right case, tracks their status, and manages the review workflow. Version control and retention policies are built in.

AI document intelligence

The AI layer extracts key information from documents and classifies them by type. It uses a combination of OCR, form recognition, and LLM-based extraction. Critically, the AI surfaces its confidence levels and routes low-confidence items to human review.

The human-in-the-loop pattern

For documents where classification or extraction matters (and in this domain, it usually does), we built a review workflow:

• 1.AI processes the document and extracts structured data.
• 2.If confidence is above threshold and the document type allows it, the extraction is accepted automatically.
• 3.If confidence is below threshold, or if the document type requires review, the item goes to a human reviewer.
• 4.The reviewer can accept, correct, or reject the AI's output.
• 5.Corrections feed back into evaluation data that helps us improve the models.

Technical stack

The platform runs on Azure, with .NET Core services for the backend API, React for the frontend, and Azure SQL for persistence. Document processing uses Azure AI Document Intelligence for OCR and form extraction, with custom LLM pipelines for classification and semantic extraction.

We use Azure Blob Storage for document storage with immutable retention policies. The audit trail is append-only and stored separately from operational data.

What we learned

The most important lesson: AI features that ship and stay shipped are the ones where you can see what the AI did and why. Observable AI workflows, structured outputs, and human approval gates aren't overhead. They're the difference between a feature that works and one that gets turned off.