ENDEع
All projects
Live in my home

Paperless AI

German paperwork filed by an LLM: title, sender, type, amount, deadline, and a short summary in Arabic.

LLMStructured outputPythonPaperless-ngxCost guard

Problem

German letters arrive constantly, and deadlines hide in long paragraphs. I wanted every document sorted and every deadline reminded, without reading each one twice.

Architecture

  1. 01 Scan / upload
  2. 02 OCR
  3. 03 LLM, no tools, JSON schema
  4. 04 Title · type · amount · deadline
  5. 05 Reminder push

Key decisions

  1. The model gets no tools and answers only through a strict JSON schema. Document text is treated as data, never as instructions.
  2. One model call per document does everything (classification, amounts, deadlines, tax relevance), which keeps cost predictable.
  3. A hard monthly budget with a shared lock: every AI feature reserves its maximum cost before calling and records the real cost after.
  4. It never overwrites values I corrected by hand.

What I learned

  • Structured output plus validation beats clever prompting.
  • Cost control has to be designed in from the first line, not added later.

Repository

Repository coming soon