Facturized
AI-powered SaaS for intelligent invoice extraction
Spanish utility bills are complex, multi-page PDFs with non-standard layouts. I built an AI-powered SaaS to convert these documents into structured JSON with production-grade reliability.
The Challenge
Utility companies in Spain use wildly different PDF formats. Traditional template-based OCR systems frequently fail when layouts shift even slightly. The core challenge was building a system that could handle these variations without manual intervention while maintaining 95%+ accuracy for financial data.
Technical Solution
Implemented a multi-stage pipeline: OCR for text layer extraction, followed by a compressed LLM (GPT-4o-mini/Claude-3-Haiku) for intelligent field mapping and semantic understanding of complex tables.
Product Support Angle
Designed a comprehensive monitoring dashboard to track extraction failures in real-time. Created L3 support flows for re-processing "orphaned" invoices and documenting API integration guides for enterprise clients.
Technical Architecture
Pipeline Flow
- 1. User Ingestion: PDFs are uploaded via REST API or Dashboard.
- 2. Intelligent OCR: Context-aware extraction of raw text layers.
- 3. LLM Refinement: Semantic parsing of fields (CUPS, Amount, Dates).
- 4. Cross-Validation: Confidence scoring and data normalization.
- 5. Society Association: Logic to link invoices to existing client accounts.
Support Engineer Insights
"Building the system was only half the battle. Supporting it meant building tools that allowed us to troubleshoot why a specific invoice failed to map and providing clients with 'API Integration Guides' that reduced onboarding time by 60%."
View API Integration Guide (Sample)