Minarai

Technical manual localization & QA

RFP demonstration

Inspect source evidence, translation risks and layout integrity before a machinery manual is allowed to become a release artefact.

Release state
Specialist review required
Live pipelineAny source language → Japanese

Run the manual-localization pipeline

Upload any born-digital or scanned PDF. Minarai detects the source language, combines native geometry with multimodal OCR and layout understanding, produces a Japanese style-preserving review draft, and blocks unsafe release.

Primary
Claude Fable 5
Independent QA
Gemini 3.1 Pro
Language routing
Source auto-detected · Japanese target
What actually runs
  1. 01
    Adaptive document understanding
    Native text where reliable; multimodal OCR for scans, diagrams and embedded artwork.
  2. 02
    Contextual translation
    Page image, source blocks and glossary are translated together; identifiers and units are protected.
  3. 03
    Style-aware reconstruction
    Typography, spacing, inline icons, callout frames, line art and diagram geometry are preserved or reflowed deliberately.
  4. 04
    Two-layer QA gate
    Coverage, token and fit checks run deterministically, followed by an independent visual reviewer.

This creates proposed ontology records with page evidence. The generated PDF is a review projection—not an automatically approved release manual.

Verified production snapshot. The Minarai side below comes from live run 03362df5 using the uploaded source PDF, Claude Fable 5 translation and independent Gemini 3.1 Pro visual QA. The customer-supplied PDF remains available only as a provenance-neutral comparison.

Real pipeline output
Input diagnosis
Hybrid PDF + VLM
Native text plus targeted raster-label OCR
Review scope
Pages 66–68
Viewing PDF page 30: Film routing phases 2 and 3
Approval
0 of 3 pages
No page has specialist sign-off
Language coverage
English → Japanese fixture
Live upload auto-detects the source language
Printed page
Compare source with
90%
Source evidence
English source · PDF page 30
Evidence linked
Source evidence: machinery manual page
Selected evidence
Verified Minarai production run
Production run 03362df5 · proposed pending specialist approval
Evidence linked
Verified Minarai production run: machinery manual page
Selected evidence
Evidence · page 68 · Automated release checks

Production pipeline found no page-level defect

The live run achieved complete text coverage, retained every protected identifier and passed independent visual QA on this page. Human terminology approval is still required.

Automated check passed
Source evidenceEN

reversal triangle · lower triangle (Ti) · infeed belt (N1)

Proposed Japanese / outputJA

反転三角板 · 下部三角板（Ti）· インフィードベルト（N1）

Required action

A machinery specialist should approve or revise the proposed terms before release.

Human gate preview

Evidence before approval

Automated extractionComplete
Translation lifecycleProposed
Specialist decisionNot recorded
Release exportHeld
This automated check passed, but it does not approve the page as a whole.
Processing trail

Native extraction when possible, OCR when necessary

The system routes each page by evidence quality, keeps every proposed translation linked to its source region, and prevents critical findings from becoming a downloadable artefact.

01

Diagnose

Complete

Hybrid document route selected

Native PDF text handles body copy; VLM-guided OCR is used only where wording is baked into diagram pixels.

02

Structure

Complete

Blocks and diagrams linked

Instructions, callouts, component identifiers and visual regions remain inspectable as evidence.

03

Translate

Proposed

Japanese draft proposed

Translation remains proposed until terminology and safety semantics receive human approval.

04

QA + review

Proposed

Automated checks passed

Visual, semantic and protected-token checks passed; machinery terminology and source safety wording still require expert sign-off.

What is demonstrated now

  • Hybrid routing: native PDF extraction plus VLM/OCR for raster labels.
  • Per-page source evidence, artefact observations and protected-token checks.
  • Critical semantic and layout findings that prevent unsafe export.
  • Synchronized document inspection with region-level highlighting.

What the customer PoC must validate

  • Representative manuals for every requested source language, including scanned or photographed pages.
  • Provider bake-off with OCR ground truth, cost and latency.
  • Customer terminology, translation memory and technical-translator approval.
  • Collision-free searchable PDF regeneration across representative page types.