Localization as an engineering discipline, not a spreadsheet.
Omniglot turns translation work into structured jobs: file ingestion and validation, translation memory, glossary enforcement, in-context QA, and guarded AI assistance — built by someone who ran localization for a living.
Great products ship
embarrassing translations.
Not for lack of translators — for lack of system: context gets lost between handoffs, terminology drifts per file, QA happens out of context, and the spreadsheet becomes the database.
Strings arrive as exports of exports; structure and placeholders break in transit.
"Save" the button and "save" the verb translate differently — spreadsheets don't know that.
The same feature gets three names across three files without an enforced glossary.
Raw LLM translation is fast, confident, and inconsistent — exactly what a brand can't afford.
Two generations,
one doctrine.
Omniglot Wrapper handles the file-heavy reality — import, validate, export without breakage. Omniglot Next Gen is the spec-driven platform: projects, keys, documents, TM, glossaries, QA, and AI assistance that operates inside guardrails.
Ingestion validates structure and placeholders on the way in; exports preserve them on the way out.
Translation memory pre-fills what's already been decided; glossaries lock the terms that must not drift.
Model selection, style rules, and glossary constraints shape the draft — the AI proposes within the system, not around it.
Segment review happens where the string lives, with flags routed to a human — SME-style checks, not blind approval.
email to translators →
paste back, pray →
broken placeholders in prod
tm + glossary pre-pass →
guarded ai draft →
in-context qa →
clean export, same shape
Source file to shipped strings,
in six stages.
Every stage is a system step with validation — the human reviews meaning, not formatting damage.
Import files; validate structure and placeholders.
Memory pre-fills previously decided translations.
Locked terms and style rules bound the draft space.
Model-selected, rule-shaped drafts per segment.
In-context review; flags route to a human decision.
Out in the same structure it came in — intact.
Real screens.
Running fully local.
Captured from local mode — SQLite on disk, local session auth, zero cloud dependencies (SPEC-020, dual-database seam). The same platform runs unchanged against Supabase.
Omniglot is localization engineering — not a spreadsheet relay, not a raw LLM pipe, not a black-box TMS.
It is what happens when localization expertise and AI-assisted product engineering are the same person.