Google’s Deepmind division has posted a case study on its website of using the Gemma family of large AI language models in the Ukrainian service «eDozvil». They are running on NVIDIA H100 GPUs in a Ukrainian data center.
Google’s Deepmind division has posted a case study on its website of using the Gemma family of large AI language models in the Ukrainian service «eDozvil». They are running on NVIDIA H100 GPUs in a Ukrainian data center.
The ePermit service is the first state service in Ukraine that uses artificial intelligence for automatic preliminary document verification, which operates on the Diya portal. It was created by specialists from the Ministry of Economy, Ecology and Agriculture of Ukraine, who integrated the Gemma 3 27B model into the workflow.
Google noted that using AI to improve the efficiency and accuracy of document processing was a logical decision, but Ukraine’s legislative framework imposed some important security restrictions. In particular, all AI systems used in Ukrainian public services must operate in a closed infrastructure, without external API calls or dependence on cloud technologies.
«We must operate within strict regulatory and security requirements, keeping all operations in a completely closed environment without external APIs or cloud technologies. Gemma has provided us with a rare balance between powerful reasoning and secure local deployment,» said CDTO and Deputy Minister of Economy, Ecology and Agriculture of Ukraine Oleksandr Tsybort.
Thanks to Gemma, the ministry has created a system that analyzes document information extracted from submitted scans and combines it with data from government registers, then recommends a decision to the government official assigned to review the application.
To decide which LLM to base their model on, the ministry’s developers tested several open-source models to find the one that best suited their tasks. The team found that Gemma 3 27B consistently outperformed the other models tested in terms of accuracy, logical consistency, and controllability, especially in structured decision-making workflows.
«Gemma’s performance and stability made it the most appropriate choice. An additional advantage of Gemma is the ability to deploy it efficiently on-site due to its relatively compact size. The Gemma model line is one of the best open models that is suitable for most cases, including the Ukrainian language and context, which is especially important as Ukraine continues to develop its own national Ukrainian LLM,» said Dmytro Voytech, AI Advisor.
ePermit uses the Qwen 2.5 7B VL model to extract text data from scanned documents, and the Gemma 3 27B to ensure the completeness of the application, consistency of information, and compliance with certain requirements, such as the minimum age and education of applicants. The model then analyzes and makes one of three decisions, which it passes on to the official responsible for reviewing the application: issue the license, reject the application, or conduct additional manual verification.
The model is currently deployed in Ukraine using VLLM and runs on NVIDIA H100 GPUs in a Ukrainian data center.
Ukrainian developers are already working to expand the model to include other types of licenses besides veterinary licenses, and are also refining it to improve understanding of the Ukrainian language and terminology specific to the documents. This will include creating a scoring system to further improve the model’s performance and accuracy.
The team is also experimenting with a unified multimodal approach, refining Gemma 3 based on a dataset of scanned documents to improve visual understanding of documents.
Recall that Google recently announced a small version of its open-source model Gemma, designed to run on local devices. According to the company, the new Gemma 3 270M can be quickly configured and maintains high performance despite its small size.
