Facturias
Multimodal invoice-processing SaaS
What it is
Structured extraction from invoices with a self-hosted vision-language model (Qwen2.5-VL) plus RAG over Spanish tax regulation to classify and validate entries, with confidence scoring, a labelled evaluation set tracking field-level precision and recall across model versions, and human review on low-confidence fields. VeriFactu compliant, multi-tenant FastAPI backend on PostgreSQL 17 with row-level security and pgvector.
Vision-language extraction
Self-hosted Qwen2.5-VL reads invoices as documents rather than as OCR text dumps, preserving table structure, line items and stamps that flat OCR loses.
RAG over Spanish tax regulation
Retrieval over the Spanish tax code classifies and validates each entry, so a deduction is justified by a retrievable rule rather than by model intuition.
Confidence and human review
Every field carries a confidence score; anything under threshold is routed to human review instead of being written silently.
Measured, not assumed
A labelled evaluation set tracks field-level precision and recall across model versions, so a model upgrade is a measurable decision.
Multi-tenant by construction
PostgreSQL 17 row-level security isolates tenants at the database, not in application code. pgvector holds the regulation embeddings.