Translation that speaks your sector's language
General-purpose translation is great for a menu and mediocre for a cholera-response field guide. We pair a configurable AI model with a translation memory of real, in-domain examples.
Free for non-commercial humanitarian & social-impact use · no account needed to start
BetaCurrently in beta — translations are actively tested and human-verified as we refine quality.
How it works
Grounded in real, in-domain examples
For every piece of text, Translate retrieves the closest human-verified examples from the domain's translation memory and hands them to the model as guidance — so it phrases things the way professionals in the field actually do.
A sentence, a paragraph, or a whole document
Full-text + dense vector search find the closest in-domain examples
Verified examples + subject-tuned instructions
Self-hosted Gemma · private · EU-resident
Right terminology, right register
WASH specialists & professional linguists — built independently of your text
The domain's curated translation memory
Getting the words right
Where generic translation slips, domain memory holds
General-purpose engines translate WASH terminology literally. Anchored in a sector corpus, Translate reaches for the accepted term instead.
Illustrative EN→FR examples. Pulled from the WASH translation memory in production.
The proof
Measurably better on domain content
We benchmarked the engine against the tools teams reach for today — DeepL, Google, Lara — and frontier LLMs, on held-out WASH test sets across 12 languages.
Automated neural metrics — for now. Scored with XCOMET-XL, a model-based quality metric, not human ratings. We're now working with native-speaking domain/sector professionals who score these outputs against human validators to ground the numbers in expert judgment. Full method on the benchmark page.
Mean score vs. the engines teams use today
XCOMET-XLZoomed axis — bars start at 82, not 0, so the gap between engines is visible.
Built to grow
One engine, many domains
WASH is the first live domain — but the domain is a first-class dimension, not a fork. New domains slot in with their own corpus, prompt, and best model per language.
Domain is a dimension, not a fork
Water, Sanitation & Hygiene — its own corpus, prompt & best model per language
Same pattern, new corpus
Same pattern, new corpus
Same pattern, new corpus
No translation memory — for off-domain text
Water, Sanitation & Hygiene — its own corpus, prompt & best model per language
Same pattern, new corpus
Same pattern, new corpus
Same pattern, new corpus
No translation memory — for off-domain text
Future-proof
Gets better as the models do
We continuously test new base models as they're released — and the moment a better self-hosted, secure model proves out on our evaluation set, we adopt it. Your translations keep improving, while your text never leaves our private, EU-resident infrastructure.
Continuously evaluated
New base models — like Gemma — are benchmarked on our domain evaluation set the moment they're released. We measure, not assume.
The best model wins
When a stronger self-hosted model proves out, we adopt it. Your translations keep getting better over time — no migration, no action needed on your end.
Privacy is never traded
Every upgrade stays self-hosted, secure, and EU-resident. Quality goes up; your data exposure stays at zero.
Why teams trust it
Private, resilient, and free for non-commercial impact
Private & EU-resident
Self-hosted models keep your text out of any big-tech AI cloud — and database and document storage stay in the EU region
Frugal & lower-impact
We choose small base models like Gemma — 26B parameters, not the trillion-plus behind ChatGPT or Gemini — so each translation uses a fraction of the energy, running in low-carbon data centres
Free for impact
Free for non-commercial humanitarian, non-profit & public-interest work
Translate your first paragraph now
No account needed to start. Text, documents, or a metered REST API — all on one shared engine.