What you receive
| Extracted field | Value | Evidence |
|---|---|---|
| Document type | Consolidated articles of association · confidence 98% | 21-page document |
| Administrator | João X. (sample document) | Page 7, clause 12 |
| Powers clause | individually and without value limit, open, operate and close bank accounts (sample document) | Page 9, clause 15 |
| Rule: powers to open an account | Compliant | Clause 15 |
| Rule: articles in force | Compliant | Filing of Nov 2024 |
| Open items | None | — |
| Desk | Awaiting ratification | — |
Real return format, in readable form, over a sample document. Full technical reference on the Developers page.
No field decides for you. The checks say what was met and where the proof is.
How the platform works
Every file goes through the same steps, in the same order, and each step records what came in, what went out and the source consulted.
- Reading. Every page is read by OCR and becomes searchable text, with the position of each passage preserved as evidence.
- Classification. The platform identifies the type of each document and the confidence of that identification. An unexpected document is flagged, not guessed.
- Extraction with evidence. The fields the operation needs are extracted with their page and passage of origin, each with a reading confidence.
- Parties and cross-checks. People and companies are related across documents; external data comes in here: company standing, who signs, liens on the property.
- Rules. Your rules are applied cell by cell; each comes out compliant, non-compliant or not assessed, always with the evidence.
- Desk and opinion. The opinion is proposed with the findings; your team confirms, corrects or rejects, and the decision is recorded.
How we measure accuracy
- Two ways for every cell. The deterministic rule and the probabilistic analysis check the same data point; they agree, it is verified; they diverge, it goes to the desk.
- Confidence per field. Below the threshold your rule sets, the cell is not decided by the platform.
- Validation before production. Every model is validated against reference sets before promotion, and accuracy is tracked per document type.
- Measured on your operation. Accuracy and turnaround numbers come from the pilot, on files representative of your process, per document type and source.
Use cases
- Signature and powers validation inside the bank’s flow. Business account opening: who may open, who may sign, which articles are in force. Answered inside the pipeline the bank already runs.
- Document analysis with evidence. Credit, collateral or formalization file classified, extracted and checked against your rules.
- Prior check of registry requirements. What the target registry requires, checked before filing, rechecked after correction.
What only this architecture does
- The missing part of the file, fetched at the source. Property record, certificates and the articles in force come straight from notaries and commercial registries; external data and the asset appraisal complete what the document does not say.
- Source per field. Every extracted data point is linked to its page and passage of origin. AI extraction becomes auditable proof.
- Articles in force. Among several filings, we identify the consolidation in force and who holds powers today.
- Requirements per registry. The target registry’s rules applied to the file before filing, curated for each project for the registries in your operation.
How it integrates
Versioned REST API, webhooks, or the GarantiaBR web dashboard for teams that operate without integration: the same files, findings, desk and trail, on a screen of their own. Shared or dedicated environment.