The national Handwritten to Data challenge, in which developers created solutions based on artificial intelligence for recognizing Ukrainian handwritten text, has ended in Kyiv. The winners' technology is planned to be integrated into government services, in particular the e-Permit system, in the near future.
The national Handwritten to Data challenge, in which developers created solutions based on artificial intelligence for recognizing Ukrainian handwritten text, has ended in Kyiv. The winners' technology is planned to be integrated into government services, in particular the e-Permit system, in the near future.
According to the Ministry of Digital Information, the competition brought together over 870 participants from different companies and countries. In total, 198 teams uploaded over 6,800 versions of AI models to the Kaggle platform.
The top ten teams advanced to the finals, presenting their products to an expert jury. The scoring was based on a combined system: 70% of the score depended on the accuracy of recognition of closed document sets, and 30% on expert assessment of the architecture and scalability.
The first place and the main prize went to the Dantists team, which included Data Scientist of the Mriya project Konstantin Zagorulko and military serviceman, ex-specialist of PrivatBank Oleksandr Dudnev. Their development combines two neural networks: one is responsible for direct text recognition, and the other checks the result and corrects typical errors. This solution turned out to be not only the most accurate on real documents, but also the most suitable for technical deployment in registries.
Second place went to Senior ML Engineer Dmytro Kovalenko of Jaroona. His system is based on a combination of several specialized detection models with a large language model-editor. Although this product requires significant computing resources, it provides maximum quality proofreading of the most complex materials.
Third place went to a developer from Japan, Tetsuya Fujimuri, who decided to donate his winnings to support Ukraine. His team focused on compactness, building the process around a small Qwen 3.5 4B language model, retrained for Ukrainian script. Such a system can be deployed on a single RTX 4090-level video card, which makes it cost-effective for deployment on large data sets.
The competition was based on RUKOPYS, Ukraine's first open annotated dataset of handwritten text , which contains materials spanning over a hundred years. This database is planned to be expanded to create new AI models that will help digitize the State Archives' holdings faster and automatically check tasks in educational applications.
