AI-powered Receipt and Expense Management Platform

What we did

• 

  Rebuilt the back-end and key parts of the app

  We rebuilt the NestJS back-end and reworked the parts of the Ionic app that carry the product’s complexity: receipt capture and processing, sign-up and login, sync, reporting, and subscriptions. Along the way we resolved the issues users were hitting — capture crashes, duplicate sales-tax entries, date-handling edge cases — set up a proper staging environment, automated builds and releases for both app stores, added error monitoring, and brought both apps up to current platform requirements (Android target API 36, iOS 26 SDK).

• 

  Improved security and infrastructure

  We introduced secure credential management and two-factor access across the project’s tooling, closed the security findings from our audit, and moved deployments from hand-managed virtual machines to containerized, repeatable releases with CI/CD. Data is encrypted in transit and at rest, and the AI layer no longer depends on an external server with a tunnel into SaveIT’s environment.

• 

  Strengthened integrations

  We reworked the QuickBooks Online integration — fixing two-way sync, adding safe fallbacks for categorization, and making it testable against a sandbox company rather than live books. On the email side, we consolidated on Gmail and Microsoft accounts via secure OAuth (retiring manual IMAP configuration) and combined authorization and inbox access into a single onboarding step, so new users understand what SaveIT needs and why before they grant it.

• 

  Redesigned reporting

  Reports were rebuilt around a clearer category breakdown with drill-down, a sales-tax toggle, and vendor logos throughout the transaction list — small changes that make a monthly expense review noticeably faster.

• 

  A new website

  Alongside the product work, our design and development team created a new website for SaveIT — a sharper hero, a clearer how-it-works story, refreshed positioning, and a fully responsive layout — now live at saveitreceipts.com.

• 

  Moved AI receipt recognition from the old setup to a modern model stack

  The original recognition pipeline combined a cloud OCR service with a custom categorization model operated outside SaveIT’s infrastructure — hard to run independently and impossible to improve.

  Rather than pick a model on reputation, we built a benchmark harness on real receipt data and evaluated several frontier models on extraction accuracy, categorization quality, and cost per thousand receipts. That harness turned accuracy from a judgment call into a number SaveIT tracks — and, more than once, it showed that a problem lived in the data rather than the model, which changed what we fixed.

  We then rebuilt the recognition engine on Anthropic’s Claude running on AWS Bedrock, inside SaveIT’s own cloud account, so the company owns and controls its AI layer. The engine ships with the safeguards a production system needs: confidence scoring, a “needs review” guard for low-confidence results, a fallback model path, and score-gated alerts when categorization quality drifts. It went live with a full database backup and end-to-end verification, and receipts that had been waiting were processed automatically once it was on.