Fatura turns photos of receipts and invoices, in Turkish, Arabic, or English, into verified, structured data and a ready-to-use Excel file.

Most AI receipt readers just trust whatever the model returns. Fatura checks the math:

1. Multimodal extraction: the image goes to Gemma 4 / Gemini through the Gemini API, which returns store, date, currency, category, line items (original + translated name), tax, and total as JSON.
2. Deterministic validation: a rules engine checks that the sum of line items equals the total (±0.05), including tax-exclusive invoices.
3. Self-correction loop: if the numbers don’t add up, the exact arithmetic error (each item, the difference, and whether something is missing or duplicated) goes back to the model, which re-examines the image. Up to 2 corrections.
4. Excel export: one formatted workbook per invoice, plus a combined workbook. Totals are live formulas, and Arabic files are laid out right-to-left.

Built for real-world use

• Automatic fallback between models (Gemini 3.8 Flash ⇄ Gemma 4 31B) when one is overloaded
• Auto-detects which features each model supports (JSON mode, thinking level)
• Retries with backoff, time budgets, and per-visitor rate limiting
• The API key stays on the server, each visitor’s session is isolated, and images are never stored
• A full audit trail of every attempt and correction

Stack: .NET 10 class library (C# records, async HttpClient), ASP.NET Core minimal API, Streamlit UI, openpyxl, and 35 offline xUnit tests. MIT licensed.

Example: a real Turkish telecom invoice with 12 line items, including negative carry-over lines and a tax-included column. Fatura extracted it correctly and validated 282.98 vs 283.00 within tolerance.ءذ