NeuroBridge
AI Enterprise SolutionsGPT & Claude

CodeNova

Who: The accounting assistant at a Baku distributor or small accounting bureau that keeps its books in 1C, handling paper and phone-photographed documents that e-qaimə doesn’t cover: supplier waybills, cash receipts, and invoices from foreign or non-VAT suppliers. What goes wrong: They retype every document into 1C by hand. Mistakes slip in: a mistyped VÖEN, total or ƏDV, the same invoice entered twice, or the wrong item card chosen because the supplier’s product name doesn’t match.

Open the demoVideoGitHub

Preliminary score

56.0/ 100Promisingaverage of 2 judges
  • Value for the user19.5 / 25

    A specific problem and a clear outcome from solving it.

  • Prototype and use of AI14.5 / 30

    A working core scenario, and what AI actually contributes to the result.

  • Quality testing7.5 / 20

    Tests, examples of failures, and comparison with the current approach.

  • Feasibility9.0 / 15

    Data requirements, running costs and a clear next step.

  • Originality5.5 / 10

    A meaningful difference from common solutions.

What the judges wrote

  • GPT-6 AstraPreliminary56/100
    Strengths
    Your project targets a concrete workflow: getting paper receipts and supplier documents into 1C for Baku accounting teams. Local VÖEN/ƏDV checks, catalog matching, and accountant review make the proposed solution more useful than standalone OCR. Your reported backend tests cover several pipeline stages and invalid inputs, and you explicitly acknowledge that real-invoice browser testing remains incomplete.
    To improve
    Demonstrate a real invoice end to end, including accountant corrections and a successful 1C import, identifying any mocked integrations. Reconcile your submission’s PaddleOCR/Odoo stack with the README’s Claude/Tesseract/SAP stack so the implemented AI contribution is clear. Test a representative, anonymized invoice set and report field accuracy, catalog-matching errors, concrete failures, and total review time versus manual entry. Validate your 0.00085 AZN/document cost estimate against the actual model, image/token usage, hosting, and retries.
  • Claude Opus 5.5Preliminary56/100
    Strengths
    Your problem framing is sharp and local. You name a specific user (the 1C accounting assistant at a Baku distributor or bureau) and concrete errors: mistyped VÖEN, ƏDV or totals, duplicate invoices, and wrong item cards. Your observation that e-qaimə covers only born-digital B2B documents, leaving paper waybills, cash receipts and non-VAT invoices to manual entry, is a credible gap. Your pipeline is well chosen for accounting: OCR extraction, VAT and arithmetic checks, duplicate detection, RapidFuzz catalog matching, ok/review routing, and a 1C-ready CSV export. Your deck also gives per-document cost and pricing figures.
    To improve
    Your materials contradict each other on the core AI. The submission says PaddleOCR, the deck says vision LLMs, and the README describes Claude, Tesseract and SAP OData. Show clearly what actually runs and what AI adds beyond OCR plus rules. Your 27 passing unit tests show the code works but not how accurate it is. Run 20–50 real Azerbaijani invoices and receipts end to end, report field-level accuracy for VÖEN, totals and line items, and show concrete failure cases. Back the 4 min vs 12 sec/doc claim with a timed comparison against manual entry, and replace the lorem-ipsum 'Tools' slide with your real stack.

The team

  • Khalida HajizadaLead

    developer

  • Zhala Suleymanova

    data