Hacki
About the project
What we built
Invoice is a tool that takes a photo of an invoice or delivery note and automatically compares it against the original purchase order, flagging discrepancies in price, quantity, or missing items — without any data ever leaving the device.
The problem
Reconciling invoices against what was ordered is manual, repetitive, and error-prone: someone has to read every line, compare prices, and spot differences by eye. For small businesses and shops, this means silent losses that nobody ever catches.
How it works
- Capture: a photo of the invoice/delivery note is taken (regardless of image
quality, angle, or lighting).
- Extraction: an AI model running entirely locally via the QVAC SDK reads the
image and structures the data (items, quantities, prices) into JSON.
- Matching: a deterministic rules engine compares the invoice against the purchase
order, with an AI fallback for ambiguous cases.
- Result: discrepancies are listed with severity and a clear explanation, ready
for a human to review in seconds.
Why local matters
All inference runs on-device through @qvac/sdk. Financial data from the invoice never leaves the machine: no cloud API calls, no per-inference cost, and it works completely offline.
Tech stack
- Inference: QVAC SDK (@qvac/sdk), model [fill in: name and quantization]
- Backend: NestJS / TypeScript
- Matching engine: deterministic rules + AI fallback for ambiguous cases
Project status
Tested against a non-cherry-picked set of invoices, including low-light and skewed-angle cases. [Fill in with accuracy % once metrics are ready]