Case study
Product · Windows desktop app
InvoicePro
Turns invoice PDFs, scans, photos and handwritten bills into checked Excel sheets, using an AI model that runs offline on the user's own computer.
The problem
Small businesses, freelancers and accounting firms still retype invoice details into Excel by hand. It's slow and error-prone. Uploading financial documents to a cloud AI service isn't acceptable for many of them, and much of what they receive is phone photos, poor scans or handwritten bills that plain OCR can't read. Nepali and Indian invoices add lakh-formatted amounts, Bikram Sambat dates and PAN/VAT numbers that generic tools get wrong.
What I built
A Windows desktop app, now on version 3.1. Drop in a batch of PDFs, photos, scans or a ZIP, and a vision-language model bundled with the installer reads each page offline, with no internet or API key. Every figure is checked, you review the result next to the original, then export a formatted Excel or CSV file. It also tracks payments and catches duplicates, and is sold with a free tier and paid plans unlocked by signed licence keys.
How it works
- 01
The installer bundles llama.cpp and a 2-billion-parameter vision model (Qwen3.5-2B). It starts in the background, uses the GPU only when that's faster than the CPU, falls back if the GPU fails and restarts itself after a crash. About 15 seconds a document with a graphics card.
- 02
The model copies values exactly as printed, into a fixed JSON schema at temperature 0, with nothing carried over between documents. Plain code then does the conversions: lakh and European number formats, date order and B.S. dates.
- 03
A fusion step compares the AI's reading, field by field, with a rule-based reader and the page's own text. A misread digit is corrected only when the printed figure makes subtotal − discount + tax + shipping equal the total exactly. Printed figures are never changed.
- 04
Anything that still doesn't add up is flagged for review. The review screen shows the original next to the fields, where each value came from and live arithmetic checks, and every edit and approval goes into an audit trail.
Key decision
The AI reads. It never gets the last word.
Earlier versions deliberately used no AI: for financial data, predictable beats clever. But rules alone couldn't read photos, handwriting or unusual layouts. So version 3 adds a small model that runs on the user's own computer, and treats its answer as one opinion to check, not the truth. If its figures don't agree with the rules, the page text and the invoice's own arithmetic, a person is asked to look instead of the app quietly guessing.
Technology
- Electron
- React
- llama.cpp
- Qwen3.5-2B vision model
- Express
- Tesseract.js
- SQLite · Prisma
- ExcelJS
Have a problem to solve?
Something like this for your business? Let's talk.
Tell me what you're trying to achieve, or what you're still doing by hand. No technical brief needed.