Case Study

BuchhaltGenie

AI-powered accounting for Austrian businesses. Custom OCR, automatic VAT compliance and multi-agent architecture. 742 sessions to production.

3.000+

Sessions

71%

OCR Improvement

30.000+

Tests

CHALLENGE

The Problem

No OCR solution reliably understands Austrian receipts. Umlauts, special formats and the nuances of Austrian VAT law make generic systems unusable for the DACH region.

SOLUTION

The Approach

Custom AI training instead of generic models. Iterative development over 742 sessions with real user feedback:

Custom OCR Vision — Specially trained vision-language model based on Qwen VL. Fine-tuned with LoRA on Austrian and German receipts. 71% better recognition than standard OCR.
VAT Compliance Guard — Automatic real-time validation against Austrian VAT law. Detects missing mandatory fields and incorrect tax rates.
Multi-Agent Architecture — OCR agent, compliance agent, categorization agent and supervisor. Parallelized processing with transparent decision chains.

SCREENSHOTS

RESULTS

What Was Achieved

A production-ready SaaS solution with 71% better OCR recognition than standard systems, automatic VAT compliance and an AI assistant that explains bookings in natural language.

Tech Stack

Next.jsReactTypeScriptSupabaseOpenAIQwen VLLoRAStripeTailwind

ENTSTEHUNG

The Making Of

BuchhaltGenie was built over 742 development sessions, evolving an initial idea into a production-ready SaaS solution. The key was custom AI training: instead of relying on generic OCR, a custom vision-language model based on Qwen VL was trained, fine-tuned with LoRA on thousands of annotated receipts from the DACH region. A multi-agent architecture coordinates specialized agents for OCR, compliance and categorization. Sophie, the RAG-based assistant, answers questions and gives proactive compliance hints.

Discuss Your Solution

I build AI solutions for specific domains. If you have a problem that generic software can't solve, let's talk.

Explore Your Solution

More Projects

EventDrop

AI-Curated Event Photo Galleries

Next.js 16SupabaseStripe+4

View Details

WalkAITalkie

AI Speech-to-Text for macOS

SwiftSwiftUIWhisperKit+3

View Details

SpireSense

AI Coach for Slay the Spire 2

Next.js 16TypeScriptSupabase+4

View Details

The Approach

Custom AI training instead of generic models. Iterative development over 742 sessions with real user feedback:

Custom OCR Vision — Specially trained vision-language model based on Qwen VL. Fine-tuned with LoRA on Austrian and German receipts. 71% better recognition than standard OCR.
VAT Compliance Guard — Automatic real-time validation against Austrian VAT law. Detects missing mandatory fields and incorrect tax rates.
Multi-Agent Architecture — OCR agent, compliance agent, categorization agent and supervisor. Parallelized processing with transparent decision chains.

The Making Of