Invoice and receipt extraction
A photo or PDF goes in. Clean, structured fields come out, in your layout, every time.
The problem
General models read invoices well on a good day and invent fields on a bad one. Your accounting software wants the same twelve fields in the same shape for every supplier. Manual re-keying costs hours and introduces errors.
Data you already have
Past invoices and receipts with the values your team already entered into your accounting system. A few hundred pairs is a good start; a thousand is better. Supplier names and amounts stay in; anything the task does not need is scrubbed.
What we train
An 8-billion-parameter open vision model, fine-tuned to read your document types and emit exactly your JSON layout: supplier, date, line items, tax, totals, your cost codes. The same approach has been shown on public receipt datasets to lift structured-output accuracy from inconsistent to reliable.
How we measure it
Field-level accuracy and JSON validity on invoices the model never saw during training, compared with the untrained base model. You see the number and the failures before you accept.
Where it runs
On one GPU in your AWS account, your Cloudera cluster or your own server. Batch mode for month-end runs or an endpoint for real-time capture.
- Who it is for
- Finance teams, bookkeeping firms, logistics and procurement operations.
- Typical timeline
- Three to five weeks from intake to handover.
- You receive
- Merged weights, adapter, quantized copy, eval report, deploy recipe, training manifest.