Invoice processing is one of the most tedious tasks in any business. I built this open source tool to eliminate it - upload a PDF, image, or forwarded email and the app extracts every critical field in seconds using a structured GPT-4o agent backed by regex fallbacks and MIME parsing.

The system handles messy, real-world invoices: scanned PDFs, low-quality images, multi-page documents, and inconsistent layouts. Azure OpenAI's structured output mode ensures the extracted data is always clean and ready to use - no hallucinations, no reformatting.

The pipeline starts before AI even touches the file. A MIME parser identifies the input type — PDF, image, or email attachment — and routes it to the correct pre-processor. PDFs are converted to page images, emails are stripped of HTML and extracted, and images are normalized for consistent OCR quality.

The pre-processed content is sent to an Azure OpenAI agent using Structured Output mode — forcing the model to return a validated JSON schema with specific fields: vendor, invoice number, date, line items, subtotal, tax, and total. This eliminates hallucination risk and makes downstream parsing deterministic.

For fields the AI misses or is uncertain about, a regex fallback layer runs pattern matching for common invoice formats — dates, currency amounts, tax ID patterns, and PO numbers. This hybrid AI + rule-based approach achieves near-perfect extraction accuracy across diverse invoice layouts.

Secure cookie sessions handle auth with no external OAuth dependencies. Extracted results are stored in Supabase PostgreSQL with user-scoped access. The app is deployed on Render with environment parity to Azure App Service, keeping both options open for enterprise deployment.

Jinja2

FastAPI

Azure OpenAI (GPT-4o)

Regex + MIME

Supabase + PostgreSQL

Render