BuchhaltGenie
From Idea to Production
The Challenge
Austrian SMBs struggle with bookkeeping daily. Scanning receipts, categorizing, posting VAT-compliant entries - a time-consuming process prone to errors.
- OCR systems fail with Austrian documents (umlauts, special formats)
- No AI understands the nuances of Austrian tax law (UStG)
- Existing solutions require manual rework for 30-40% of receipts
- Integration with existing accounting systems is complex
There was no solution that addressed these problems holistically. So I built one.
The Approach
After 100+ prototypes and countless iterations, a clear approach crystallized:
Instead of relying on generic models, I trained custom models on Austrian receipts and VAT requirements.
Qwen VL as the base, fine-tuned with LoRA on thousands of annotated receipts from the DACH region.
742 development sessions with continuous feedback and improvements.
Specialized agents for OCR, compliance, categorization, and user interaction.
Key Features Built
Each feature was perfected through dozens of iterations
Conversational AI with access to the entire knowledge base. Answers questions, explains bookings, helps with compliance.
- Natural language interaction in German
- Context-aware memory
- Proactive compliance hints
Specially trained vision-language model for Austrian and German receipts with highest accuracy.
Compared to standard OCR solutions on Austrian receipts
Automatic verification of VAT compliance. Detects missing mandatory information, incorrect tax rates, and potential risks.
- Real-time VAT validation
- Automatic categorization
- Input tax optimization
Specialized AI agents work together: OCR agent, compliance agent, categorization agent, and supervisor.
- Parallelized processing
- Error recovery mechanisms
- Transparent decision chains
Features in Detail
Smart customer categorization with AI
AI-powered receipt recognition (99% accuracy)
Fully automated VAT integration
From 58% to 99% recognition rate
Automatic categorization and suggestions
Live synchronization with Austrian tax authority
Results & Metrics
Measurable improvements over existing solutions
Development Sessions
OCR Improvement
Tested Prototypes
Production-Ready
Tech Stack
Modern technologies for maximum performance and scalability
Learnings
What Worked
- LoRA fine-tuning enables fast iterations without compromising quality
- Multi-agent systems scale better than monolithic approaches
- Intensive collaboration with real users from day 1
What Didn't Work
- H100 GPUs didn't deliver better results than A100 for this use case - just higher costs
- Too early optimization for edge cases delayed the MVP by weeks
- First RAG implementation was too complex - simplicity wins
- Custom training beats generic models in specific domains
- Iterative development with real feedback is irreplaceable
- The right architecture decision early in the project saves months later
Similar Challenges?
I help companies bring complex AI projects from idea to production. If you're facing similar challenges, let's talk.