Portfolio project · document automation

Doc2Sheet

Turn invoices and receipts (PDFs, scans, phone photos) into clean spreadsheet rows. A vision LLM reads each document, and deterministic checks re-do the arithmetic, so a wrong total gets flagged instead of slipping into your books.

View the code on GitHub Run it yourself

Real results on four sample documents

Pick a document to see what was extracted and what the checks said. All four samples are fictional.

How it works

One model call per document, surrounded by plain code that doesn't trust the model.

LoadPDF pages are rendered to images; a digital PDF's text layer is kept too.

ReadAny OpenAI-compatible vision model returns JSON against a strict schema.

ValidatePydantic parses money and dates (1.234,50 € · 01.09.2026). Malformed output gets one repair round.

CheckLine maths, items vs subtotal, subtotal − discount + charges + tax = total, dates, currency.

ExportColour-coded Excel (Documents · Line items · Issues), CSV and JSON.

What the checks catch

Each document ends up ok, review or failed.

total_mismatch

Subtotal, discount, charges and tax don't add up to the printed total. Tax-inclusive pricing is recognised.

line_math · items_vs_subtotal

A quantity × unit price that doesn't match its line, or lines that don't sum to the subtotal.

due_before_issue · future_date

Impossible dates. Ambiguous ones like 05/03/2026 are refused rather than guessed.

missing_* · invalid_currency

No total, vendor, date or ISO currency code, so a person should take a look.

Run it yourself

Web UI, batch CLI and Python API. Works with Hugging Face Inference Providers, OpenAI, Gemini, OpenRouter, or fully offline with Ollama.

git clone 
cd doc2sheet
pip install -e ".[app]"
# put HF_TOKEN=hf_... in .env (or point DOC2SHEET_BASE_URL at Ollama)
python app.py                                  # web UI on http://127.0.0.1:7860
python -m doc2sheet invoices/ -o results.xlsx  # a whole folder → one workbook

Hosting a live Gradio demo on Hugging Face now requires a paid plan, so this page shows recorded results instead.