All work
052025SCAD

Arabic ↔ English Statistical Report Translator

Bilingual statistical reports in 4 hours instead of 3 days — with zero terminology drift.

Before

Bilingual reports needed 2–3 days of manual translation, and statistical terminology drifted between the Arabic and English versions — the same concept rendered three different ways across a single publication.

After

A GPT-4 translation copilot with a SCAD-specific glossary and a side-by-side reviewer cut turnaround to 4 hours, locked 200+ recurring terms to a single agreed rendering, and is now used for every quarterly release.

Challenge

Bilingual reports required 2–3 days of manual translation with statistical terminology drift between Arabic and English versions.

Approach

Built a GPT-4 translation copilot with a SCAD-specific glossary (UN SDG terms, demographic taxonomies) and a side-by-side reviewer UI.

How it was built

  1. 1

    Glossary first

    Phase 1

    Built the SCAD-specific glossary (UN SDG terms, demographic taxonomies) before any translation code — because terminology consistency, not raw translation, was the real problem.

  2. 2

    Translation copilot

    Phase 2

    Wrapped GPT-4 with the glossary as enforced context so agreed terms are used verbatim, and built a side-by-side reviewer so a human owns the final wording.

  3. 3

    Adoption

    Phase 3

    Rolled it into the publications team's quarterly workflow. Human-in-the-loop review kept trust high while the speedup did the convincing.

Key architecture decisions

Glossary-enforced translation over raw model output

Why · Free translation drifts on domain terms. Injecting an agreed glossary is what eliminated drift across 200+ recurring statistical terms.

Side-by-side human review, not full automation

Why · Published government statistics can't ship on model confidence alone. A reviewer UI made the speedup safe to adopt.

Impact

  • Cut bilingual report turnaround from 3 days to 4 hours
  • Eliminated terminology drift on 200+ recurring statistical terms
  • Adopted by the publications team for every quarterly release
18×
speedup
200+ terms
glossary
100%
adoption

What I'd tell someone building this

  • 01 · For domain translation, terminology consistency matters more than fluency.
  • 02 · A glossary is cheaper and more reliable than fine-tuning for locking terms.
  • 03 · Keep a human on the final wording when the output is published under an institution's name.

Tech stack

GPT-4Azure OpenAICustom GlossaryAngular

Ask anything about Arabic ↔ English Statistical Report Translator

AI scoped to this project · GPT-OSS 120B