Reading documents is easy. Trusting the answer is the hard part.

A document-centric intelligence engine that extracts and validates — bilingual Indonesian and English, 41 document types, every field checked against a real per-type rule rather than handed back raw.

Not just OCR

Extraction is the start, not the deliverable.

Each uploaded document runs its own pipeline — extraction, then validation — and lands on a terminal status: PASS, REVIEW or FAIL. There is no cross-document scoring layer to second-guess; the document is the unit of work.

Both languages are load-bearing, not an afterthought. Field labels, date formats and currency conventions are read bilingually because the real documents mix them on the same page.

The pipeline

Four stages, one path per document.

01 · Classify

Every page routed to the right type

PaddleOCR with fuzzy-keyword and regex anchoring sorts each page, with a hosted vision-language model as the fallback tier when the layout is ambiguous.

02 · Extract

Anchors where layout is fixed, VLM where it isn't

Government forms with a stable layout are read anchor-first from OCR and the native text layer. Invoices, contracts and medical documents — where every issuer differs — are read VLM-first.

03 · Ground

Extracted values checked against the page

Per-field OCR grounding and arithmetic reconciliation catch hallucinated values before they reach validation, on the document types where OCR is reliable enough to do it.

04 · Validate & review

A terminal status per document

Real per-type validation rules resolve each document to PASS, REVIEW or FAIL. Anything low-confidence or failed is routed to a human review queue for correction.

41 types · 9 categories

What it reads today.

Personal Identity

KTP
SIM
NPWP
Kartu Keluarga

Business Identity

NIB
Profil AHU
SK Kemenkumham
Akta (Pendirian/Perubahan)
Beneficial Ownership
Proof of Address
Surat Kuasa

Banking

Bank Statement
Loan Agreement
Financial Statement
SLIK iDeb

Insurance

Medical Resume
Medical Bill
Insurance Card
Claim Form
Medical Check-Up

Commerce

Invoice

Funding

Proof of Transfer
Purchase Order
Faktur Pajak
Bukti Potong PPh
Payslip
Surat Setoran Pajak
Payment Request
Bank Account Details

Lending

Surat Jaminan Fidusia
Sertifikat Properti
Surat Persetujuan Pasangan
Akta Pemberian Hak Tanggungan (APHT)
Surat Kuasa Membebankan Hak Tanggungan (SKMHT)

Supply Chain

Delivery Order (Surat Jalan)
Warranty Letter (Surat Jaminan Garansi)

CBAM

Electricity Bill (Tagihan Listrik PLN)
Management System Certificate (ISO / IATF)
Berita Acara Serah Terima (Handover Minutes)
Weighbridge Ticket (Slip Timbangan)
Certificate of Origin (SKA)

Measured, not claimed

Field accuracy against real documents.

Every number below is field-level accuracy on a corpus of real fixtures, published in the system design docs. The weaker types are listed at their real number rather than smoothed over — Medical Resume at 73.7% is tracked openly as the current floor. Types without a published measurement are supported but not shown here.

Insurance Card	100%	15/15 fields · 5 real fixtures
Kartu Keluarga	98.2%	331/337 fields · 4 real fixtures
SK Kemenkumham	98.3%	226/230 fields · 10 real fixtures
Akta (Pendirian/Perubahan)	100%	168/168 fields · 10 real fixtures
Beneficial Ownership	97.4%	76/78 field-instances · 8 real fixtures
Profil AHU	27/28	scalar fields exact · 7 real fixtures
Bank Statement	99.1%	multi-page transaction recall
Proof of Address	89.2%	83/93 field-instances · 13 real fixtures
Medical Bill	85.6%	154/180 fields · 5 real fixtures
Claim Form	81.2%	130/160 fields · 5 real fixtures · first handwritten type
Medical Resume	73.7%	210/285 fields · 5 real fixtures · tracked openly as the weakest type
MCU	65.0%	scalar fields · structured fields 24/24 shaped correctly

Extraction requires an account.

The processing pipeline sits behind the login gate. Sign in to upload a document and see its extraction and validation result.