Rebuilding a fragmented translation memory and making it stick (anonymised)

Background

Several years of work across vendors had left the translation memory in pieces: duplicates, conflicting terms, and mismatched language codes. Teams wanted a clean base they could trust without pausing delivery.

Stakeholders: a client‑side localisation PM, vendor PMs, translators/reviewers for core EU and APAC locales, and engineering who expected the same shape back every week.

What I built

A consolidation pipeline that aligned bilingual sources, deduplicated segments, normalised punctuation and spacing, and reconciled language tags into a single TMX with clear provenance. A compact termbase captured product names and do‑not‑translate entries with context.

Simple governance: change logs, light approval for term updates, and a QA pass that flagged brand and units issues automatically.

How we ran it together

We migrated low‑risk content first, verified that quality held, then folded in the rest in waves. Translators saw fewer fuzzy mismatches; PMs spent less time adjudicating repeat questions.

Outcome

Higher leverage from past work, faster turnarounds, and clearer decisions on terminology. The key win was not the tooling—though that helped—but the simple habits that kept the TM healthy after the clean‑up.

What I’d keep today

Keep the termbase small and visible, automate the obvious checks, and record once why a choice was made so it doesn’t resurface every quarter.