Receiptflow

Overview

Receiptflow helps accountants and businesses eliminate manual receipt entry. Users upload photos or PDFs, and the platform extracts structured financial data, categorises spend, and connects to tools like accounting software and bank feeds. I owned the backend end to end — from secure upload and cloud storage, through the security and AI extraction workers, to persisting structured results in the database.

What I built

On Receiptflow, I helped build a production document-management platform that turns uploaded receipts and invoices into structured financial data. The main challenge was designing a pipeline that could handle varied file formats, untrusted uploads, and inconsistent AI extraction quality — all at scale for multiple organisations. I solved this by architecting a staged async processing system: immediate file validation on upload, background security scanning (OCR + malware + content analysis), then AI extraction with fallback providers and automatic recovery of missing fields. For performance-sensitive areas like PDF report generation, I introduced concurrency limits and shared resource pooling. The result was a reliable, secure, and observable document-processing system integrated with banking (GoCardless), accounting (Xero), and real-time client updates.

Outcome

Receiptflow went from prototype to production. We improved extraction reliability by combining AI confidence scoring and automatic field recovery with a manual review workflow for edge cases — such as faded thermal receipts — so uncertain results are corrected by a human before they’re treated as final.

Key highlights

• End-to-end receipt pipeline: upload → storage → workers → Postgres
• BullMQ + Redis for async security and AI extraction
• Gemini extraction with retries, fallbacks, and field recovery
• Manual review queue for low-confidence results
• Upload-time validation + async malware/content scanning
• Multi-tenant RBAC for practice and business orgs
• Stripe + GoCardless billing with webhook sync
• WebSocket updates during document processing