Manual invoice and receipt reading slows down follow-up work
Invoices and receipts often require the same fields again and again: taxable supply date, document number, price without VAT and price including VAT. When these values are typed manually, people spend time reading the document, copying fields and checking typos.
The problem is not only text recognition. The useful output has to be structured so it can feed a spreadsheet, internal system, import script or another automation.
Local web app with model selection and JSON output
I built a local web tool with a browser interface and a Node.js server. A user can drag a PDF or photo of an invoice or receipt directly into the app. The app shows the document preview, so the extracted fields can be compared with the original immediately.
PDF files are read with pdftotext first. For structured extraction, the app uses local Ollama models such as qwen2.5. For images, it can use vision models such as moondream. The user can choose an available local model in the interface and save it as the default.
Recognized values appear in a form for review. After correction, the user can save JSON, copy JSON or store the corrected example for future recognition improvements.
Faster document data preparation for downstream systems
The tool speeds up invoice and receipt reading because the user does not have to manually search for every field in every document. AI prepares a draft and a human checks or corrects it where needed.
JSON output is practical for connecting the result to spreadsheets, internal administration, import scripts or additional automations. The project shows how a small internal AI tool can reduce repetitive administrative work without building a large accounting platform.
