📄 AI Invoice Data Extractor

AI-powered pipeline that extracts, validates, and stores structured data from PDF invoices — built for German small businesses to automate manual bookkeeping while enforcing German VAT compliance rules.

🔗 Live App: invoice-data-extractor-ai.streamlit.app 💻 GitHub: github.com/Akash-045/invoice-data-extractor-ai

---

Overview

Small businesses lose hours manually re-typing invoice data into spreadsheets and accounting software. This project automates that process end-to-end:

PDF invoice → text extraction → AI structured extraction → validation → database storage → dashboard

Instead of a rigid, template-specific parser, the pipeline uses the Claude API to read invoices the way a human bookkeeper would — generalizing across different vendor layouts — while a deterministic validation layer double-checks the AI's output against real German invoicing rules (correct VAT math, valid VAT rates, valid dates) before anything is trusted or stored.

Why this project

Most "AI chatbot" portfolio projects show that someone can call an API. This project instead tackles a real, unglamorous business pain point — bookkeeping — and treats AI output the way a production system should: verify before you trust it. It also reflects specific German business context: VAT rates (19%/7%), invoice requirements under §14 UStG, and the ongoing shift toward mandatory e-invoicing (XRechnung/ZUGFeRD from 2027–2028).

Architecture

PDF Invoice
    │
    ▼
[1] Text Extraction (pdfplumber)
    │
    ▼
[2] Structured Extraction (Claude API → JSON)
    │
    ▼
[3] Validation (VAT math, date checks, business rules)
    │
    ▼
[4] Storage (SQLite)
    │
    ▼
[5] Streamlit Dashboard (upload, view, filter, export)

Tech Stack

Layer	Tools
PDF parsing	pdfplumber
AI extraction	Claude API (anthropic)
Validation	Python (custom business rules)
Storage	SQLite
Data handling	pandas
Web app	Streamlit
Secrets management	python-dotenv (local) / Streamlit Secrets (cloud)

Project Structure

invoice-data-extractor-ai/
├── 01_extract_text.ipynb      # PDF → raw text
├── 02_extract_fields.ipynb    # raw text → structured JSON (Claude API)
├── 03_validate.ipynb          # validate extracted data against business rules
├── 04_store_data.ipynb        # store validated data in SQLite
├── app.py                     # Streamlit web app (full pipeline, live)
├── requirements.txt
├── data/
│   ├── raw_invoices/           # sample PDF invoices
│   └── processed/
│       ├── raw_text/           # extracted text
│       ├── extracted_json/     # structured AI output
│       ├── validated/          # validated records with status
│       └── invoices.db         # SQLite database
└── README.md

Key Features

• ✅ AI-powered extraction — generalizes across different invoice layouts without custom templates
• ✅ Built-in validation layer — catches VAT math errors, invalid dates, and unusual rates before data is trusted
• ✅ Full audit trail — every invoice keeps its validation status, errors, and warnings
• ✅ German VAT compliance logic — checks against real §14 UStG-relevant rules (19%/7% VAT rates, VAT math, totals)
• ✅ Live, interactive dashboard — upload invoices, see KPIs, filter by status, export to CSV
• ✅ Secure key handling — no hardcoded API keys; uses .env locally and Streamlit Secrets in production

Validation Rules

The pipeline doesn't blindly trust AI output. Each invoice is checked for:

• Required fields present (vendor, invoice number, date, amounts, currency)
• VAT rate is a valid German rate (0%, 7%, 19%)
• VAT amount math is correct (net × vat_rate/100 ≈ vat_amount)
• Total math is correct (net + vat ≈ total)
• Invoice date is valid and not in the future
• Amounts are non-negative Invoices are classified as Passed, Warning, or Failed — so a human only needs to review flagged exceptions, not every invoice.

Running Locally

# Clone the repo
git clone https://github.com/Akash-045/invoice-data-extractor-ai.git
cd invoice-data-extractor-ai
 
# Install dependencies
pip install -r requirements.txt
 
# Add your Anthropic API key
echo "ANTHROPIC_API_KEY=your-key-here" > .env
 
# Run the app
streamlit run app.py

Future Improvements

• OCR support for scanned invoices (not just digital PDFs)
• Support for XRechnung/ZUGFeRD e-invoice formats, ahead of Germany's 2027–2028 mandatory e-invoicing rollout
• Multi-currency and multi-country VAT rule sets
• Duplicate invoice detection
• Export directly to DATEV-compatible format for accountants

Author

Akash Samantray Data Analyst | Ironhack Berlin Data Analytics Bootcamp LinkedIn · GitHub