FormGuide — AI-assisted pre-check for Wohngeld applications

Short tagline: Local, privacy-preserving AI that pre-checks income-proof PDFs and extracts key applicant data.

Inspiration

Applying for Wohngeld is often confusing and time-consuming, especially when applicants aren’t sure whether their income documents meet submission standards. We built FormGuide to reduce that uncertainty and give applicants a quick, private pre-check, all running locally on their machine.

What it does

FormGuide converts uploaded income-proof PDFs to text, verifies whether the document is a valid income proof, and extracts essential fields such as last name, first name, and monthly income. The entire pipeline runs locally: GPU-accelerated PDF processing, a local Gemma 3 (27B) model via Ollama for validation and extraction, and a lightweight Flask UI/API for uploads and results.

How we built it

PDF processing: Docling with CUDA for fast, GPU-accelerated PDF → text conversion.
AI validation & extraction: A LangChain agent using ChatOllama (Gemma 3 27B) evaluates document validity and extracts the required fields via prompt engineering.
Frontend: Minimal Flask app exposing an upload UI and a JSON /upload endpoint.
Local-first: Ollama serves the model locally so no cloud data leaves the user's machine.

Challenges we ran into

Government-issued documents vary widely in layout and formatting. Building robust validation and consistent extraction across real-world, messy documents was the biggest engineering challenge.

Accomplishments we’re proud of

A fully local AI pipeline with no cloud dependency.
A practical tool that reduces bureaucratic friction while keeping humans in the loop for final checks.
A working prototype delivered within hackathon constraints.