Basque Government's EJIE builds high-speed AI translation tool on Red Hat OpenShift
EJIE (Basque Government) · Spain
EJIE, the Basque Government's IT management arm, built the Itzuli neural translator to translate between Spanish, French, English and Basque, replacing an older grammar-rule-based tool that struggled with long texts. Using Red Hat OpenShift to run GPU-backed containers for the neural translation models, with support from Red Hat Consulting and a Technical Account Manager, EJIE launched general and specialist legal/administrative versions of the tool. Usage grew 300% in three years: from about 2 million quarterly translation requests in early 2020 to over 6.7 million per quarter by August 2023, with peaks over 260,000 queries in a single day and over 30,000 daily users; official-service translation requests fell 40% over the same period as citizens used the self-service tool directly.
Overview
EJIE, the Basque Government's IT management arm, built the Itzuli neural translator to translate between Spanish, French, English and Basque, replacing an older grammar-rule-based tool that struggled with long texts. Using Red Hat OpenShift to run GPU-backed containers for the neural translation models, with support from Red Hat Consulting and a Technical Account Manager, EJIE launched general and specialist legal/administrative versions of the tool. Usage grew 300% in three years: from about 2 million quarterly translation requests in early 2020 to over 6.7 million per quarter by August 2023, with peaks over 260,000 queries in a single day and over 30,000 daily users; official-service translation requests fell 40% over the same period as citizens used the self-service tool directly.
The challenge
EJIE's earlier automatic translation tool, developed in 2014, used grammar-rule-based technology; it worked well with short sentences but made mistakes in long texts. EJIE needed a high-quality, automated tool supported by modern AI capabilities to translate Spanish, French and English into Basque and vice versa, and had to build new technical capabilities in GPUs and Kubernetes containers to do so.
The solution
EJIE built the Itzuli neural translator using Red Hat OpenShift to run GPU-backed containers for the neural translation models, aided by the GPU operator that simplified GPU integration in OpenShift, with Red Hat Consulting and a Red Hat Technical Account Manager supporting the project. EJIE launched a general version of the tool for colloquial language plus a more specialized version for legal and administrative tasks.
Reported business value
Usage grew 300% in three years: from about 2 million quarterly translation requests in early 2020 (with peaks of 90,000 in a single day) to over 6.7 million per quarter by August 2023 (peaking at over 260,000 queries in a single day). Over 30,000 people now use the translator daily, and over 60,000 people have installed the app on a mobile device. Official-service translation requests fell 40% over the same period as citizens used the self-service tool directly.
Sources
Open any source and check the claim yourself — that is the point of the register.
Other government & public sector entries in the register.
Estonia rolls out Bürokratt, an AI-guided virtual assistant network for public services
Bürokratt is a network of chatbots deployed on Estonian public sector institutions' websites, letting people obtain information from institutions and use public and information services via virtual assistants. It is a state-created, AI-based digital assistant that helps institutions deliver modern, efficient, around-the-clock customer service using large language models.
Austrian Academy of Sciences unlocks Ancient Greek with Mistral
The Austrian Academy of Sciences (OeAW), together with its Austrian Archaeological Institute, partnered with Mistral and services partner Reply to build Apollo, described as the first advanced large language model for Ancient Greek. Apollo is trained on a specialized corpus of 600 million words of historical Greek text plus tens of thousands of published inscriptions and papyri, helping researchers reconstruct damaged texts and identify thematic connections across collections. The OeAW reports Apollo turns work that once took years into hours, addressing over one million unread Greek papyri worldwide, with future phases planned for semantic search and handwritten inscription decipherment.
MITRE's Federal AI Sandbox accelerates government AI research with NVIDIA DGX SuperPOD
Nonprofit MITRE built the Federal AI Sandbox, powered by NVIDIA DGX SuperPOD, to give federal sponsors an affordable, centralized space to test and deploy AI and machine learning across domains such as weather forecasting, cybersecurity, and public benefits administration. The DGX SuperPOD delivers a 300-fold performance increase over MITRE's previous AI computing capabilities and supports thousands of researchers; MITRE is using it with NVIDIA Omniverse and NVIDIA Earth-2 to develop 1-kilometer precision weather forecasts with NOAA and the National Weather Service, plus a foundational model for cybersecurity threat analysis across 190 nations.
Israel's Meteorological Service Predicts the Weather Using NVIDIA Earth-2
The Israel Meteorological Service integrated an AI weather model called F3, powered by NVIDIA Earth-2 and NVIDIA PhysicsNeMo (including the CorrDiff neural network), into its forecasting system to achieve hyperlocal 2.5-kilometer resolution precipitation forecasts. F3 achieved 90 percent lower computational costs compared to running a classic non-AI numerical weather prediction model on a CPU cluster, delivering forecasts in minutes rather than hours, and saved 20-30 percent of the Tel Aviv-Yafo Municipality's preparedness budget for things like urban flood safety. CorrDiff was trained to emulate IMS's regional high-resolution ICON model using low-resolution ECMWF IFS input, enabling four high-resolution forecasts per day. Partners including the Israeli Police, the National Fire and Rescue Authority, the Ministry of Transportation and the National Road Company have adopted the forecasts, and IMS used F3 to issue a high-resolution flood map during a January storm that dropped 150 millimeters of rain in six hours.
Was this helpful?
Your feedback helps us improve our use case database