Ukrainian startup Knowledgator has created a platform that extracts structure from unstructured data: from scientific articles, financial reports, logs, or even legal documents. It is already used by large pharmaceutical companies, banks, and technology startups.
Ukrainian startup Knowledgator has created a platform that extracts structure from unstructured data: from scientific articles, financial reports, logs, or even legal documents. It is already used by large pharmaceutical companies, banks, and technology startups.
Knowledgator is one of the few Ukrainian startups that focuses not just on content generation, but on the structural understanding of texts. Its technologies help transform the chaos of unstructured data into organized and easy-to-analyze information. dev.ua spoke with Knowledgator CEO Igor Stepanov about how it works, which businesses are already using the platform, and what makes it unique.
«We train the model not just to find keywords, but to extract connections, essence, structure — to actually read the text like a human,» he says.
Knowledgator, at its core, is a universal platform for processing information. The user enters text (or an array of documents), as well as a template that describes what entities need to be found. This can be anything: company names, transaction dates, mentions of side effects, etc. The user receives structured data in the form of a table or JSON.
«Our models work with a zero-shot query: there is no need to retrain for each new task. For example, if a user needs to extract specific proteins from scientific texts, the model will cope with this without additional training,» explains Stepanov.
Today, Knowledgator is used in several industries:
«One of our advantages is the combination of zero-shot extraction and the possibility of further training. If the client wants even higher accuracy, we provide an interface for quick fine-tuning,» adds Stepanov.
According to the CEO, the accuracy of the models depends on the case: for general tasks, even in basic mode, the accuracy is high, and in critical tasks (such as processing medical documents) it cannot be less than 99%.
«Accuracy is key. That’s why we either use zero-shot or train specialized models, but always on high-quality data,» Stepanov emphasizes.
The platform is available in several formats. Basic open-source models can be downloaded from Hugging Face. More powerful solutions are available via API, and in the near future the startup will open a dashboard that will simplify access and allow you to customize the models for yourself.
The Knowledgator team consists of 12 people, mostly from Ukraine. The startup received a pre-seed round from Startup Wise Guys, grants from the Ukrainian Startup Fund, and support from Google for Startups.
The team is currently working on multimodality: in addition to texts, the platform will soon be able to work with audio and images. Support for over 100 languages, including Ukrainian, is also being actively improved.
«Our goal is to become leaders in the field of information extraction. These are not just models — this is an infrastructure that will help global companies work with chaotic data in a structured way,» concludes the CEO of Knowledgator.
We also published an article about how Ukrainian startups first got a chance to compete for funding from the European Commission separately from their European competitors. In July, the EU announced a new competition focused exclusively on deeptech teams from Ukraine that already have an MVP and the potential to scale.
