Ongkrong Consulting
  • 01Claude setup
  • 02About
  • 03How it works
  • 04Case studies
Start the diagnosis→
← Case studies
CS/02Case study · AI document translation operations

Ampor Hub — AI document translation operations

How Ongkrong Consulting designed and built a controlled AI translation operating system for a visa, legal, and official document business — replacing manual file-by-file translation while retaining every document's context, format, and structure, with humans kept in the approval seat.

An account of Ongkrong's AI document-translation engagement for Ampor Translation — covering product design, AI pipeline engineering, OCR, structured quality evaluation, human review, layout mirroring for official documents, client operations, invoicing, portal delivery, production hosting, and a companion Telegram AI receptionist. Prepared by Ongkrong Consulting · accurate as of June 2026.

Engagement at a glanceProduction-system shape
10Workstreams delivered end-to-end
12Language registry · configurable hub model
7Quality-evaluation dimensions per segment
4 wksHandoff to production

At a glance

Client
Ampor Translation — a translation and consultancy business specialising in visa, official, and legal document translation
The problem
Human translators processed every visa application, official form, legal contract, and certificate by hand — files arriving by chat, email, and walk-in; no OCR; no structured quality review; jobs, invoices, and client records in separate places
Engagement type
End-to-end: product architecture → AI translation pipeline → OCR → multi-language support → structured quality evaluation → layout mirroring → client management → invoicing → portal delivery → production deployment → AI receptionist
Quality evaluation
7-dimension Quality Assessment Framework (QAF) — protected-pattern integrity, script integrity, OCR confidence, semantic fidelity, terminology compliance, format fidelity, glossary coverage — applied per segment, per job
Document formats
DOCX · digital PDFs · scanned PDFs · JPG/PNG images · multi-page browser scans — the real formats a translation business receives, not only clean files
Language support
12-language registry with a configurable language-hub rule — every job involves one of the configured hub languages on one side, keeping the operational centre of gravity aligned to the markets served
AI receptionist
Companion Telegram bot — FAQs, service guidance, appointment booking, document upload triage, language switching, staff handoff, owner notifications, and RAG from a controlled company knowledge base
Production status
App, database, background worker, deployment path, runbooks, security baseline, and a working product surface for staff — production-system shape

The client and the problem

Ampor Translation is a translation and consultancy business. It does not only receive clean Word documents. Clients send visa applications, official forms, legal contracts, certificates, scans, phone photos, low-quality PDFs, and mixed-format material. A useful system had to work with the messy front door of the business, not only with ideal inputs. The operational requirement was broader than translation accuracy — before Ampor Hub, the workflow was fragmented at every step.

  • ▸Files arrived through chat, email, phone photos, and walk-in requests — no single intake point.
  • ▸Scanned documents required manual reading or separate OCR tools outside the workflow.
  • ▸Translation output often lost document structure — tables, headings, field positions.
  • ▸Important details — dates, passport numbers, legal references, names, IDs — could be damaged with no automated preservation checks.
  • ▸Review happened outside the system, making status and quality hard to track.
  • ▸Client records, job history, invoices, and delivery were spread across separate tools with no linkage.
  • ▸Customer questions and booking requests relied on manual response time with no capture.
The opportunity

Replace human translators for visa applications, official documents, and legal contracts with a controlled AI-assisted workflow — while retaining every document's context, structure, and format. Not just translation: translation operations — intake, OCR, AI pipeline, quality evaluation, human review, layout mirroring, client records, invoices, portal delivery, and customer intake in one system.

Our role — ten workstreams

Ongkrong Consulting worked across product, engineering, AI workflow design, deployment, security, and operational handoff — ten workstreams from first intake to customer support, designed as an integrated system where each feeds the next.

WS 01

Translation Operating System

Internal command centre — job flow, status tracking, worker pipeline, review, delivery.

WS 02

Document Intake

DOCX, digital PDF, scanned PDF, image, browser scan — real formats, automatic routing.

WS 03

Multi-language Controls

12-language registry, language-hub rule, glossary, templates, protected patterns.

WS 04

OCR & Quality Review

AI vision OCR, document classification, 7-dimension QAF, segment approval flow.

WS 05

Layout Mirroring

Layout analysis, region mapping, layout editor, final DOCX/PDF export.

WS 06

Client Operations

Client records, job history, notes, services catalogue, invoices, VAT, deposits.

WS 07

Client Portal

Token-gated portal — clients view and download their own jobs and invoices.

WS 08

Production Architecture

Single container, Postgres, HTTPS, background worker, health checks, runbooks.

WS 09

AI Receptionist

Telegram bot — FAQs, booking, document triage, RAG, staff handoff, notifications.

WS 10

Security & Governance

Auth, role gating, upload validation, portal token model, secret handling, runbooks.

Main job flow — upload to deliveryTracked pipeline · staff-facing
1 · IntakeStaff uploadDocument + client + translation direction
2 · PipelineWorker runsExtract → OCR → classify → translate → assemble
3 · QAF7-dim evalConfidence scores, flags, hard blocks per segment
4 · DeliveryReview → approveStaff review, edit, approve · DOCX/PDF · portal link
Ongkrong-built pipelineStaff-facing interface
Document intelligence & routingAutomatic · no staff pre-sort
Input typeDetectionProcessing pathPipeline entry
DOCXFile extension + MIME type
Structural XML parser → Document tree extractor → Section analyser
Structured segments
Digital PDFText layer present
Position-aware text extractor → Region grouper → Section analyser
Positioned segments
Scanned PDFNo extractable text layer
Page renderer → AI Vision OCR per page → Text assembler
OCR segments
JPG / PNG imageImage MIME type
Direct AI Vision OCR → Text assembler
OCR segments
Browser scanStaff-initiated capture
Multi-page PDF export → Scanned PDF path
OCR segments
Document classification → translation registerRuns before every job
ClassDetection signalsTranslation registerControls activated
LegalLegal terminology, clause structures, section numberingHigh formality — precise legal register
Name + reference protection, glossary enforcement, strict fidelity
OfficialGovernment headers, seal markers, form-field patternsFormal — government / administrative register
Date + ID preservation, place-name and entity protection, layout-mirror path
FormField labels, blank fields, table-grid structureLiteral / field-mapped translation
Field mapping, exact value preservation, position-aware output
ReportSection headers, paragraph structure, numbered listsProfessional — section-aware
Heading translation, section integrity check, structure validation
GeneralDefault — no strong structural signal detectedStandard — neutral register
Basic date, ID, and code protection only

The human stays in control

AI accelerates the work; staff still own the final output. No translated document reaches a client without human approval — and the QAF tells staff exactly where to look first. Before translation the system analyses and classifies each document; after translation the 7-dimension Quality Assessment Framework evaluates every segment before it reaches the review surface.

Quality Assessment Framework — per-segment evaluationEvery segment · every job
#DimensionWhat is measuredSignalThreshold
01
Protected-pattern integrity
Dates, IDs, codes, emails, reference numbers detected in source, verified in output
Hard blockAny mutation detected
02
Script integrity
Target-language output validated for correct Unicode range and script encoding
Hard blockInvalid characters found
03
OCR confidence
Character-recognition score assigned per segment by AI Vision OCR
Auto-warnScore < 0.80
04
Semantic fidelity
Word-count ratio between source and translation — extreme deviation signals truncation or hallucination
Auto-warn> 1.6× or < 0.55×
05
Terminology compliance
Glossary term-match rate — required terms checked against the active job glossary
Auto-warn< 95% term match
06
Format fidelity
Structural element count — headings, tables, lists, field labels matched source to output
Review flagAny count mismatch
07
Glossary coverage
% of job-specific terms present in the active glossary before translation
Advisory< 80% coverage
Hard blockAuto-warnReview / advisory
7 dimensions · 3 signal types
Segment confidence distributionBy design tier
High · score ≥ 0.85
71%
All dimensions pass · auto-approve eligible
Medium · 0.65–0.84
22%
Review required · individual review
Low · score < 0.65
7%
Flagged for re-check · re-OCR or manual entry
All clear

Approve-high-confidence workflow

All 7 dimensions pass · confidence ≥ 0.85 · no flags. Batch-approve eligible segments without individual review.

Flags present

Full segment-by-segment review

One or more dimensions flagged. Staff must review each segment individually before the job can be approved.

Hard block

Segment rejected · not deliverable

Protected pattern mutated or script integrity failed. Segment must be manually corrected. Delivery blocked until resolved.

Documents that look like documents — and the back office around them

Many translation jobs are not plain text. Official documents need layout sensitivity: forms, certificates, tables, headings, seals, margins, and field positions. The system analyses the original layout, maps translated content back into regions by section and field position, lets staff adjust in an editing surface, and exports a DOCX or PDF suitable for client delivery. For official documents, layout fidelity is not cosmetic — it is the delivery standard the client expects. Around that, Ampor Hub also carries the business operations: client records and job history, a services catalogue, and invoices in USD and KHR with deposits, discounts, and VAT — every invoice linked to its job, every job to its client. No separate spreadsheet, no separate invoicing tool, no separate file store.

Language coverage & continuous improvement

The platform supports a 12-language registry with a configurable language-hub rule: every job must involve one of the configured hub languages on one side. That lets the business handle multilingual demand while keeping the operational centre of gravity aligned to the markets it serves. The hub model governs which pairs are valid; review-driven feedback makes every subsequent job smarter.

12-language registry · configurable language-hub modelEvery job includes ≥ 1 hub language
Pairing typeValidityRouting
Hub A ↔ Hub B✓ Supported — primary hub bridge
Direct translation · highest-priority pair
Hub A ↔ Other✓ Supported — hub-A spoke
Direct translation with hub-A-side controls
Hub B ↔ Other✓ Supported — hub-B spoke
Direct translation with target-script integrity guardrails
Other ↔ Other– Not in scope
Must involve a hub language — server-side validation prevents invalid pairs

How the system learns from every job

  1. 01

    Job intake

    Document enters; pipeline classifies and extracts content. Active glossary, templates, and patterns applied.

  2. 02

    AI translation

    Segments translated using the current knowledge base. QAF evaluates every segment before review.

  3. 03

    Human review + edit

    Staff correct errors, edit segments, flag issues. Every edit is a signal the system captures.

  4. 04

    Corrections captured

    Terminology errors, missed patterns, register issues, and glossary gaps identified and logged.

  5. 05

    Knowledge base updated

    Glossary extended, protected patterns refined, template instructions improved.

  6. 06

    Every future job benefits

    Updated controls applied to all subsequent translations — compounding accuracy over time.

Delivery, reliability, and a serious security posture

A controlled delivery channel

Clients receive a token-gated portal link to view their own jobs, download files, and see invoices — scoped to one client, isolated from staff routes, admin-managed tokens. No ad-hoc file sharing, no email attachments, no ambiguity about which version is final.

Small footprint, not fragile

Single web container with HTTPS ingress and health checks; internal Postgres with database-backed job queueing and atomic claiming; an in-process background worker with heartbeat monitoring and restart handling; migration runner at boot, plus backup, restore, and handoff runbooks.

The operating system reaches the front door

A companion Telegram AI receptionist handles FAQs and price guidance from a controlled knowledge base (RAG), appointment booking, document-upload triage, language switching, and staff/owner handoff with context — so staff receive a warm, documented enquiry, not a cold one.

Security treated as part of the product

Staff authentication with admin and staff roles and server-side route gating; strong session-secret enforcement, password hashing, and login throttling; portal token isolation; upload validation and size limits; backup and secret-handling runbooks documented for handoff.

From handoff to production in four weeksWeeks · Ongkrong-managed
Track
1234
Cloud environment
Database migration
Application deployment
Worker + job queue
HTTPS + health checks
Security hardening
Staff onboarding
AI receptionist deploy
UAT + QAF validation
Production sign-off
Ongkrong-managed deliveryClient participation at weekly gates

Deliverables

Major artifacts produced across the ten workstreams — from translation pipeline to production operations.

  1. 01Production web application for translation operations — job dashboard, status tracking, and staff workflow
  2. 02DOCX, digital PDF, scanned PDF, image, and browser-scan intake pipeline with automatic routing
  3. 03AI Vision OCR pipeline — per-page recognition for scanned PDFs, phone photos, and browser scans
  4. 04Document classification system — legal, official, form, report, general — with register-aware translation
  5. 057-dimension Quality Assessment Framework (QAF) — applied per segment, per job, before human review
  6. 06Segment confidence scoring — high / medium / low tiers driving quality-gate decisions
  7. 0712-language registry with configurable language-hub rule and server-side pair validation
  8. 08Script-integrity guardrails, protected-pattern handling, place-name and entity protection, document-type terminology controls
  9. 09Glossary and template controls — accumulated and refined through the continuous-improvement loop
  10. 10Layout editor and final DOCX/PDF export path for formatted official documents
  11. 11Client records, job history, and notes — linked throughout the platform
  12. 12Invoice builder — services catalogue, USD/KHR, VAT, deposits, discounts, PDF preview
  13. 13Token-gated client portal — view, download, invoices, scoped to one client per token
  14. 14Background worker — atomic job claiming, heartbeat, restart handling, recovery strategies
  15. 15Production deployment path, backup and restore runbooks, security and handoff documentation
  16. 16Telegram AI receptionist — RAG knowledge base, FAQs, booking, document triage, staff handoff, owner notifications

What changed for the client

Before
  • —Files arriving through chat, email, and walk-in — no single intake point
  • —Scanned documents required separate OCR tools
  • —No automated quality checks — errors caught only by human re-read
  • —Translation output losing document structure
  • —Review status and quality difficult to track
  • —Client records, jobs, and invoices in separate tools
  • —Customer questions handled manually, one at a time
  • —Delivery by email or ad-hoc file share
After
  • ▸AI handles visa, contract, and official document translation — humans review and approve
  • ▸OCR support for scans, phone photos, and browser scans
  • ▸7-dimension QAF flagging issues before human review begins
  • ▸Multi-language support anchored around the configured hub languages
  • ▸Layout mirroring for formatted official documents
  • ▸Client records tied to job history and invoices in one system
  • ▸Invoice generation inside the same workflow
  • ▸Token-gated portal for controlled client delivery
  • ▸AI receptionist for questions, booking, and lead capture
  • ▸Continuous-improvement loop compounding accuracy over time

What made this engagement different

A quality framework, not just a translation prompt.

Seven evaluation dimensions, three signal types, and three gate outcomes — applied to every segment of every job. Quality is measured continuously through the pipeline, not reviewed only at the end.

Built around the messy front door, not the ideal demo.

Five document formats, automatic classification into five document types, and an OCR path that handles the real inputs a translation business receives — not only clean Word documents.

AI output connected to operations that compound.

The continuous-improvement loop feeds review corrections back into glossary, protected patterns, and templates — so each visa, contract, and official document makes the next one better.

App showcase — see it in action

A short walkthrough of the live Ampor Hub translation operating system — intake, AI pipeline, quality review, and delivery.

Ampor Hub app showcase video thumbnailWatch the demoYouTube ↗
Work with Ongkrong

If your business processes visa applications, official documents, or legal contracts through manual translators, we can build you a controlled AI translation operation that retains document context, format, and structure while keeping humans in the approval seat. That is the brief Ampor Hub was built against.

Start the diagnosis→How the process works

Prepared by Ongkrong Consulting. Accurate as of June 2026.

Ongkrong ConsultingMelbourne · AU

Review the process. Build the control. Leave clear ownership.

Start the diagnosis→
Pages
  • Claude setup
  • About
  • How it works
  • FAQ
  • Glossary
  • Start the diagnosis
Company
  • Contact
  • Privacy
  • Terms
© 2026 Ongkrong ConsultingABN 70 543 844 199