TruelineTurning document piles into clean data, with a human in the loop

Role

Design, development, deployment (solo)

Timeline

1 day (~5 hours), solo

The problem

Small businesses receive invoices and receipts as PDFs and phone photos. Someone has to retype every one of them into a spreadsheet.

It is slow, nobody enjoys it, and the mistakes only surface later — in the accounts, where they cost real money.

The approach

• Upload a document and it comes back as structured data: supplier, dates, line items, tax and totals. Check it, then export a batch to CSV.
• AI extraction is never completely right, so the review step is the product rather than an afterthought. The extracted fields sit beside the original document, editable, and the app checks the sums itself: when the line items do not add up to the printed subtotal, it says so in plain words and shows the exact difference.
• The key decision was to tell the AI to copy what the document says and never to correct it. That sounds backwards until you follow it through — if the AI quietly fixes a wrong total, nobody ever learns the document was wrong. Copying it exactly is what lets the arithmetic check catch it.
• Only genuinely uncertain fields are flagged, because if everything is highlighted people stop looking. Nothing is ever approved automatically, and nothing the AI returns is saved until it has been checked.
• Deliberately not built: teams and roles, billing, accounting integrations, a mobile app. One idea done properly is worth more than five half-finished ones, and knowing where to stop is part of the job.

The result

• •Tested against the real Claude API on four sample documents: 6–7 seconds and roughly two US cents per single-page document.
• •The arithmetic check does its job on a document built to break it — the AI reported the printed subtotal of £1,420.00 while the line items came to £1,240.00, and the app flagged the £180.00 gap instead of silently choosing one. An unclear date (03/04/2026) is left blank and flagged rather than guessed.
• •170 automated checks across four suites, all passing: the extraction pipeline, sign-in, every API route, and a full upload-to-export run.
• •Honest limits: this is a demo built to show the approach. Four sample documents demonstrate the workflow rather than an accuracy benchmark, and the cost figure is measured on single-page files — a dense multi-page scan will cost more.