AI that reads every receipt so nobody has to type them

The challenge

Expense admin is manual by default: someone opens each receipt, reads the vendor, date, total and tax, decides a category, and types it into a spreadsheet. It scales linearly with volume, it is error-prone at exactly the moments accuracy matters most, and the resulting data lands too late to be useful for any decision.

What we did

1. 01

   Built an extraction pipeline that takes a PDF or image, renders and pre-processes it, then runs an LLM pass that returns structured fields — vendor, date, totals, tax, line items and a category — rather than raw OCR text.

2. 02

   Added a workflow engine of twelve automations that fire after every extraction: auto-categorisation, duplicate detection, missing-field checks, currency conversion at live rates, budget tracking, expense alerts and scheduled exports.

3. 03

   Wired first-class integrations to the tools the data actually needs to reach — Gmail for alerts and summaries, Google Sheets for auto-exported rows, Google Drive for document backup, and Slack for instant notifications.

4. 04

   Designed the full product surface: a marketing site, authentication, a dashboard with spend analytics, a searchable results library, a template system for recurring document shapes, and integration and settings screens.

The outcome

Delivered end to end to the client. Uploading a document now produces structured, categorised, validated data with no manual entry — and the workflow layer means that data reaches the spreadsheet, the inbox or the Slack channel on its own, rather than waiting on somebody to move it.