AI Invoice Agent

The AI-Powered CFO for Danish SMEs. A "Human-in-the-Loop" auditing tool that extracts data from invoices, validates Danish CVR numbers, handles split-VAT logic for grocery receipts, and flags anomalies before they hit your accounting software.

Built with Pydantic AI for structured LLM extraction with tool calling, FastAPI for the REST API, and React for the frontend.

Deployed app: https://invoice-auditor-production.up.railway.app/

Quick Start

This project uses uv for Python package management.

1. Install Dependencies

uv sync

2. Setup Configuration

Create a .env file in the root directory with your OpenAI API key:

OPENAI_API_KEY="your_key_here"

3. Run the App

Backend API:

uv run uvicorn invoice_auditor.api.app:app --reload --port 8000

API docs available at http://localhost:8000/docs.

React frontend:

cd frontend
npm install
npm run dev

Opens at http://localhost:5173. Requires the backend running on port 8000.

Architecture

Browser
  └─► React frontend          Drag-and-drop upload, result display
        └─► FastAPI (REST)     Receives image, returns structured Invoice JSON
              └─► Image preprocessing      Normalise to standardised JPEG
                    └─► Pydantic AI agent  LLM extraction with two tools:
                          ├─ lookup_vat    Checks line items against Danish VAT rules
                          └─ validate_cvr  Validates vendor CVR via the Danish registry
                                └─► Deterministic post-audit
                                      VAT math verification, currency conversion,
                                      and Green / Review / Red status assignment

The key design split: the LLM agent handles the inherently fuzzy work (reading handwriting, parsing unstructured layouts, inferring missing fields). Everything that can be expressed as a rule - VAT arithmetic, CVR lookup, status thresholds - runs as plain Python in the post-audit layer, keeping it testable and reliable.

API Endpoints

Method	Path	Purpose
GET	/health	Health check
POST	/api/v1/audits	Upload invoice image, run audit
GET	/api/v1/audits	List audits (filter by status, paginate with limit/offset)
GET	/api/v1/audits/{audit_id}	Get a single audit by ID

How It Works

1. Image preprocessing - uploaded images are converted to standardized JPEG
2. Pydantic AI agent - a GPT-5-mini agent extracts structured data from the image, calling tools:
   • lookup_vat - checks each line item against Danish VAT rules
   • validate_cvr - validates vendor CVR numbers against the Danish business registry
3. Deterministic post-processing - VAT math verification, currency handling, and status assignment run as plain Python (not LLM) for reliability
4. Result - the invoice is classified as Green (auto-approved), Review, or Red (issues found)

Disclaimer

This is an MVP. Data is processed locally or via API. Ensure you comply with GDPR and the Danish Bookkeeping Act when handling real financial data.