Computer Vision & OCR Auto Document Scanning

Module 1 · Document AI

Computer Vision & OCR Auto Document Scanning

Automated extraction and authenticity verification for Indonesian merchant legal documents during onboarding.

RFI module deep-dive

Computer Vision & OCR Auto Document Scanning

Prototype page exposes input contracts, processing stages, structured outputs, quality controls, privacy controls, integration points, PoC questions and evidence needed before production acceptance.

Open PoC Benchmark Lab

Inputs

Document typesNIB, SIUP/OSS, Akta, NPWP; extensible allowlist

File formatsPDF, JPG, PNG; configurable size/quality thresholds

IdentifiersCompany / document IDs; masked where not required

Processing modeSynchronous API or batch ingestion

Processing Pipeline

Pre-flight

Malware scan, MIME validation, image quality check and secure temporary object creation.

Classification

Document-type classifier identifies NIB / NPWP / Akta / permit format and version.

OCR & layout understanding

Extract text, key-value relationships, tables and signature/stamp zones.

Authenticity signals

Analyze recapture artifacts, editing inconsistencies and visual manipulation signals.

Cross-document matching

Normalize names/addresses/IDs and compare consistency across documents.

HITL routing

Low-confidence fields and authenticity alerts create review tasks.

Structured Outputs

document_typeNIB / NPWP / Akta / Permit

fieldsStructured key-value data with field confidence

authenticitytamper / recapture / suspicious pattern signals

consistencyCross-document match score and mismatch reasons

reviewHITL required + reason codes

evidenceDocument hash, model version and processing timestamp

✓

Automatic document classificationRecognize Indonesian company documents such as NIB, SIUP/OSS, Akta and NPWP.

✓

Structured field extractionExtract document number, company name, address, KBLI/LOB, management structure, NPWP and relevant fields.

✓

Authenticity / manipulation detectionDetect digital tampering, recaptured screenshots, photo-of-photo and other suspicious manipulation.

✓

Cross-document consistencyCompare company/person names and identifiers across NIB, Akta, NPWP and other documents.

✓

Field-level confidenceReturn confidence per extracted field rather than one opaque score.

✓

Human-in-the-loopRoute low-confidence or high-risk fields to manual review.

✓

Indonesia document supportModel/versioning process must support changed Indonesian document formats.

#	Question to provider	Prototype status	Evidence / response expected
1	Which Indonesian official documents are natively supported and what OCR accuracy is measured for each?	PoC response	Provide per-document precision/recall or field accuracy on labeled Indonesian samples.
2	Which methods detect document authenticity/manipulation and what accuracy / FAR-FRR is observed?	PoC response	Provide test methodology, attack classes and error rates.
3	Is the model specifically trained for Indonesian documents and how are format changes handled?	PoC response	Provide training/validation approach and update governance.
4	What is the output format, per-field confidence and human-review workflow?	PoC response	Provide JSON schema, confidence semantics and review thresholds.
5	What is latency per document and maximum throughput?	PoC response	Provide p50/p95/p99 latency and sustained/concurrent throughput.
6	Which deployment options are supported and where are documents processed/stored?	PoC response	Provide SaaS/private/on-prem architecture and data residency map.

Data minimization

Extract and retain only fields required for KYB decisioning; suppress unrelated document content.

Sensitive-document isolation

Raw scans are stored separately from analyst-facing structured fields; access is more restrictive.

No secondary training

Vendor/model training on merchant documents is disabled unless separately assessed and contractually approved.

Retention & deletion

Raw document retention is shorter than derived KYB records where legally/operationally feasible; deletion propagates to caches and backups.

Illustrative structured output

Schema is intentionally explicit to support decision-engine integration, explainability and audit. Values are simulated.

{ "module": "document_ai", "document_type": "NIB", "fields": { "company_name": {"value": "PT Contoh", "confidence": 0.993}, "nib": {"value": "********1234", "confidence": 0.997}, "kbli": [{"code": "4791", "confidence": 0.981}] }, "authenticity": {"risk": 0.12, "signals": []}, "cross_document": {"score": 0.96, "mismatches": []}, "human_review": false }

Document Investigation Workbench

Interactive upload/preview, structured fields, authenticity indicators and cross-document comparison.

Open Workbench →