Document AI
Invoices, contracts and forms turned into clean, validated data.
OCR and LLM extraction, validation rules, confidence scores and a human review queue — with export straight into your ERP or accounting system. Romanian and English documents.
DOCUMENT AI, DONE PROPERLY
Less retyping. More checking what matters.
Documents come in, structured data goes out — and a person sees only the fields that need a second look.
Most companies still have people copying numbers from PDFs, scans and photos into an ERP or a spreadsheet: supplier invoices, receipts, delivery notes, signed contracts, application forms. It is slow, it is error-prone, and it does not scale when volume grows. Classic OCR templates help only as long as every supplier uses the same layout, which they never do.
We build document pipelines that combine OCR with large language models, so extraction works across layouts and languages, and then wrap them in what production actually needs: validation rules, confidence scores, a review queue for uncertain fields, an audit trail and a reliable export into your systems. Our team has been building and running production software in Cluj-Napoca since 2013.
Extraction that reads like a person
OCR turns scans and photos into text with positions; an LLM maps it to your schema: supplier, VAT ID, dates, line items, totals, contract parties and clauses. Native PDFs and e-Factura XML are read directly.
Validation and human review
Totals must add up, VAT IDs must be valid, the supplier must exist in your ERP. Every field gets a confidence score; anything uncertain goes to a review queue with the original page next to it.
Classification, routing, export
Documents are classified on arrival and sent to the right flow or team. Approved data goes into your ERP or accounting system through its API or import format, with a link back to the source file.Documents we work with. Supplier and customer invoices, receipts and expense reports, delivery notes and goods received notes, purchase orders, bank statements, contracts and annexes, application and onboarding forms, and ID documents. Inputs can arrive by email, upload, scanner folder, API or e-Factura, in Romanian or English.
Contracts, not just invoices. For contracts we extract parties, dates, value, duration, renewal and termination terms, penalties and liability clauses, and flag clauses that differ from your standard wording. A lawyer still decides; the system makes sure nothing is missed in a long annex. For questions across many contracts at once, this pairs well with a RAG knowledge assistant.
Personal data and sensitive documents. ID cards, contracts and HR forms contain personal data, so we design for GDPR from the start: only the fields you actually need are extracted, sensitive values can be masked in logs and review screens, retention is limited and access is role-based. When documents must not leave your control, the whole pipeline can run as private AI on your servers or on our own cloud and GPU cloud in Romania, with open-weight models instead of external APIs.
Stack. OCR engines such as Tesseract or cloud OCR where allowed, vision-capable LLMs (OpenAI, Anthropic Claude, or open-weight models like Llama, Mistral and Qwen served with vLLM), Python services, PostgreSQL for documents, fields and audit trail, and queues for volume. When a document should trigger multi-step work, such as matching an invoice to an order and a delivery note, we build it as one of our AI agents.
- Extraction schema for your document types and fields
- OCR and LLM pipeline for scans, photos, PDFs and e-Factura XML
- Validation rules: totals, VAT, dates, master data checks
- Confidence scores per field and adjustable thresholds
- Review queue with the source page and highlighted fields
- Classification and routing by type, supplier or department
- Contract clause extraction and deviation flags
- Export to ERP and accounting through API or import files
- Audit trail, accuracy reports and GDPR-aware data handling
- Private deployment on your servers or our Romanian cloud
HOW WE WORK
From a folder of documents to data you can trust
Measured on your real documents before it touches your ERP.
Sample and scope
You share a representative set of documents, anonymised if needed. Together we decide which fields matter, where they go and what a wrong value would cost.
Schema and rules
We define the extraction schema, the validation rules and the master data checks, plus the confidence thresholds that send a field to review.
Pilot and measure
We run the pipeline on your sample and report field-level accuracy, review rate and failure cases, so you decide on facts, not on a demo.
Review queue
Your team gets a simple screen to confirm or correct uncertain fields. Corrections are logged and feed back into prompts, rules and templates.
Integrate and go live
We connect the inputs and the export to your ERP or accounting system, test on a copy, then switch on in production with monitoring.
Operate and improve
We watch accuracy and review rates over time, add new document types and suppliers, and adjust thresholds as confidence grows.Most of the work is on the integration side: matching suppliers, mapping accounts, handling duplicates and credit notes. If your systems need to be connected first, see our ERP integration work. For how a project like this can be priced, see our engagement models.
FAQ
Questions we usually get
How accurate is the extraction?
It depends on your documents: scan quality, layout variety, handwriting, language. That is why we start with a sample of your real documents and measure field-level accuracy before anything goes live. Every field carries a confidence score, and fields below the threshold you choose go to a person instead of straight into your ERP.
Can the documents stay on our own infrastructure?
Yes. For contracts, ID documents or anything sensitive we can run OCR and open-weight models on your servers or in a private environment on our own cloud and GPU cloud in Romania, so documents are not sent to external AI APIs. Where a hosted model is acceptable, we can use one with data retention settings agreed with you.
Do you work with Romanian documents and e-Factura?
Yes. We process documents in Romanian and English, including diacritics, Romanian VAT codes (CUI), dates and number formats. e-Factura XML can be used as an input, where the structured data is read directly instead of guessed from a PDF, or as an output when a flow needs to produce it.
Which ERP or accounting systems can you export to?
Any system with an API, a database we are allowed to write to, or a file import format. That covers platforms such as SAP, Microsoft Dynamics, Odoo and local Romanian accounting software, as well as custom systems. We map the extracted fields to your chart of accounts, suppliers and cost centres, and we test the export on a copy before it touches production data.
What happens when the model is not sure?
The document does not get pushed through anyway. Low-confidence fields and failed validation rules send it to a review queue, where a person sees the original page with the field highlighted, corrects or confirms the value and approves it. Those corrections are logged and used to improve prompts, rules and templates over time.
Drowning in documents? Talk to an engineer.
Send us a few sample documents and tell us where the data should end up. We will tell you honestly what can be automated and what should stay with a person.