📄 AI Invoice Data Extractor

The Problem: Small businesses spend 15-20 hours/week manually entering data from unstructured, messy PDFs.

The Solution: This app automatically parses PDFs, forces a Large Language Model (via OpenRouter) to extract the data into a strict JSON schema, and mathematically validates the totals before exporting to CSV.

🚀 Features

• Structured AI Output: Uses pydantic and the openai SDK to force the LLM to return clean, predictable JSON data (no conversational hallucinations).
• The "Hallucination Moat": A dedicated validator checks the AI's math (Subtotal + Tax == Total). If the math fails, the invoice is flagged for human review.
• Separation of Concerns: Modular backend architecture (parser.py, ai_extract.py, validator.py) separating the dumb text extraction from the smart business logic.
• Safety Guardrails: Hard limits on PDF page counts to prevent API token limit crashes and cost overruns.
• Streamlit Dashboard: A clean, drag-and-drop web UI for end-users to process batches of invoices and download CSV reports.

🛠️ Architecture

extractor/
├── parser.py           # Uses pdfplumber to rip raw text
├── ai_extract.py       # Pydantic schema + OpenRouter(OpenAI SDK) API call
├── validator.py        # Business logic & math validation
└── exporter.py         # Converts extracted JSON to Pandas DataFrame -> CSV

💻 How to Run Locally

1. Clone the repository
2. Create a virtual environment: python3 -m venv .venv
3. Activate the environment: source .venv/bin/activate
4. Install dependencies: pip install -r requirements.txt
5. Add your OpenRouter API key to a .env file (OPENROUTER_API_KEY=your_key)
6. Run the app: streamlit run app.py

🔮 Roadmap (V2 Enterprise Upgrades)

• Async Processing: Implement asyncio to batch process 50+ invoices concurrently.
• Network Retries: Implement tenacity for exponential backoff if the AI API times out.
• Structured Logging: Add server-side logging for production monitoring.