Loading page...Загружаем страницу...正在加载页面...
AtlasRepo

A local AI that types up a Brazilian accountant's receipts (and knows when it's wrong)

Description and use

A fully local AI system processes ~100 Brazilian accounting documents weekly, achieving 95 % field accuracy and reducing manual entry to about 5 minutes per week.

Problem

A self‑employed accountant in Brazil receives ~100 documents a week from small‑business clients and currently types each one into a spreadsheet by hand.

Outcome

The system achieves 95% field accuracy on 10 fictitious documents, with all problematic documents routed to review. Average processing time is 2.2 s per text PDF and 4.0 s per image, totaling about 5 minutes for a week’s worth of documents.

Workflow

  • Drop files into a Streamlit page and select the accounting month.
  • Each document is read by an open‑weight model running locally in Ollama.
  • Python validates the model output and flags rows that need human review.
  • The accountant edits flagged rows in an editable table and exports the result to Excel (entries sheet and summary by category).

Evidence

A self‑employed accountant in Brazil gets ~100 documents a week from her small‑business clients: NF‑e invoices (PDF), NFC‑e receipts photographed on a phone, Pix transfer screenshots, and the occasional handwritten recibo.

Source

Each document is read by an open‑weight model running locally in Ollama.

Source

Python validates everything the model returned and flags the rows a human must check.

Source

95% of fields correct. All three misses come from the blurry receipt.

Source

Average time on an AMD RX 9060 XT (Ollama, Vulkan): 2.2 s per text PDF, 4.0 s per image. That's about 5 minutes for a week of documents.

Source

  • Record date: 2026-10-16T01:41:19.222Z