Back to Overview
Case Study #3Document Verification EngineVision Pipeline & Verification Engine Architecture

Onboarding Form Verifier (OCR)

From 1 business day to 4 hours: automated document verification returning ~18 daily team hours.

Tesseract OCRVision LLMFastAPIReact 19n8nPostgreSQL

Simulated example: a field can appear on several pages of the paper form. Every occurrence is checked against the digital record. The digital record is the source of truth, so a page that differs needs its form revised. Here the account number differs on page 2, the company name differs on pages 1 and 3, and the admin phone is empty on page 6. The team's daily verification effort goes from ~20 team hours / day to ~2 team hours / day, and branch turnaround from 1 business day to ~4 hours.

IMPACT SIMULATION · V-3Simulated example

Form Verifier

The same value appears in several places: on different pages of the paper form and in the digital record. Every occurrence is checked against the digital record, and one page that disagrees is enough to flag the field.

  1. Read
  3. Cross-check
  5. Score
  7. Review
One value, many places3 of 50 fields need a human
  • Main account numberRevise p. 2
    Page 10123-4567-89
    Page 20123-4567-98
    Page 30123-4567-89
    Page 60123-4567-89
    Digital record0123-4567-89
  • Company nameRevise p. 1, p. 3
    Page 1PT Maju Sentausa
    Page 3PT Maju Sentausa
    Digital recordPT Maju Sentosa
  • Admin phoneEmpty on p. 6
    Page 10812-0000-1234
    Page 6empty
    Digital record0812-0000-1234
  • Tax numberMatches everywhere
    Page 100.111.222.3-444.000
    Page 300.111.222.3-444.000
    Digital record00.111.222.3-444.000
  • Company addressMatches everywhere
    Page 1Jl. Contoh No. 12
    Page 3Jl. Contoh No. 12
    Digital recordJl. Contoh No. 12
  • Admin emailMatches everywhere
    Page 1admin@contoh.co.id
    Page 6admin@contoh.co.id
    Digital recordadmin@contoh.co.id

+ 44 more fields match in every place

Confidence score
88%Needs review
507595
RejectReviewApprove

The officer decides. The system only recommends.

Verification team, 3 people × 8 hoursmeasured
Before~20 team hours / day
After~2 team hours / day
verification workreturned to the team
Branch turnaroundmeasured
~4 hours1 business day

Business Impact

Figures labeled 'projected' or 'estimated' reflect business model targets rather than realized metrics.
measured
1 business day→~4 hours

Branch verification turnaround SLA

Average turnaround time for branch submission processing

measured
~20 team hours / day→~2 team hours / day

Daily manual verification effort for the 3-person team

Returned ~18 of 24 collective daily working hours (±90% reduction)

derived
Baseline→±380 team hours / mo

Cumulative monthly team capacity returned

Calculated across ~21 business days for the 3-person verification unit

measured
Manual comparison→6 templates · 50 params

Verification scope handled

287 automated field mapping rules

The Problem

When opening accounts for corporate clients, branch officers submit digital applications while clients fill out physical paper forms. Head office verification teams were tasked with manually verifying both documents line-by-line across dozens of fields per form (~50).

A 3-person team spent roughly 20 of their combined 24 daily working hours reviewing 80 dossiers a day. This manual bottleneck stretched branch turnaround SLA to an entire business day and increased the risk of missed discrepancies.

The Solution

We developed an automated multi-tier verification engine:

  1. Document Digitization: Scanned paperwork passes through vision OCR and layout-aware text extraction.
  2. Field Comparison Across Every Page & Alias Matching: A field can appear on several pages of the paper form. Every occurrence is checked against the digital record, using alias matching for spelling variations, and one page that disagrees is enough to flag the field.
  3. Multi-State Field Classification: Every field is classified into one of five states: Exact Match, Minor Typo, Partial Match, Mismatch, or Empty.
  4. Weighted Confidence Scoring: The system calculates an aggregate confidence score to provide actionable recommendations:
    • ≥95%: Recommend approve
    • 75–94%: Needs review
    • <75%: Recommend reject
  5. Exception-Only Review Interface: The UI collapses all matching fields by default, guiding reviewers to the flagged fields only.

Business Impact

Verifying onboarding dossiers manually across 6 form templates consumed ~20 of the 3-person verification team's 24 daily working hours, stretching branch turnaround SLA to a full business day. Automating field-level comparison and alias matching cut branch turnaround to an average of ~4 hours and gave the team back ~18 of their 24 collective daily working hours (~380 team hours per month across ~21 business days, derived), allowing reviewers to focus exclusively on flagged discrepancies with a complete audit trail.

Engineering Principle

Show the Reviewer Only What Matters

The goal of automated verification is not to eliminate human oversight, but to eliminate human fatigue. By automatically confirming identical fields and isolating exceptions (for example, 47 matching fields and 3 with discrepancies), reviewer focus remains sharp on genuine compliance risks.

Behind the Decision

Behind the Decision

Rather than passing entire documents to an end-to-end LLM for opaque verification, we decoupled extraction from comparison: OCR and vision models extract structured text, while a deterministic comparison engine evaluates field-level alias matching for spelling variations with an explicit audit trail.

Interested in shipping similar systems?

I partner with teams to design, evaluate, and ship production-grade AI systems with verified operational impact.

Let's talk on LinkedIn