Transcript of the interview for Ukrainian Radio.

Broadcast recording from Ukrainian Radio

Valentyna Troyan: Ukrainian Radio is live on air, Valentyna Troyan is in the studio. Today we will talk about the creation of an innovative knowledge base called "Wiki: Crimea is Ukraine" — a tool for systematizing and promoting the true history of Ukrainian Crimea based on hundreds of concepts processed by a language model. Our guest is Andrii Bogdanovych, founder of digital projects, innovation and AI expert, and developer of the digital knowledge base "Wiki: Crimea is Ukraine". He is on the line with us. Greetings, Mr. Andrii!

Andrii Bogdanovych: Hello!

Valentyna Troyan: First, let's tell our listeners about the "Crimea is Ukraine" platform itself, in case someone doesn't know, and then we will talk about its transformation into a Wiki.

Andrii Bogdanovych: More than six years ago, my colleagues and I launched a project where we decided to collect all reliable information about Crimea. This includes scientific and publicistic works, archival materials, and everything connecting Crimea informationally with mainland Ukraine: about the occupation of Crimea, earlier periods, resettlement, and deportation of various peoples from Crimea. The main idea was to create a repository for such information.

Recently, we decided to diversify these materials because we noticed that most visitors were only those specifically interested in archival and academic materials. But obviously, there was a lot of content on the website that could be presented differently and made more engaging for a broader audience. Therefore, we decided to implement it as a wiki knowledge base.

Valentyna Troyan: What tools did you use? How long did this work take? And you know, when we talk about artificial intelligence and its application in any process nowadays, the immediate question arises: to what extent can we trust the resulting product?

Andrii Bogdanovych: To build this knowledge base, we decided to use a pattern called LLM Wiki. It was proposed by Andrej Karpathy, one of the co-founders of OpenAI. In fact, he worked not only at OpenAI; he is a scientist who has been involved in AI research for a long time. He proposed a concept drawing an analogy to software engineering: the LLM agent acts as a developer; the knowledge base we form is like the codebase, and Markdown editors used to view these materials (such as Obsidian) serve as the IDE.

How does it work? We take existing materials. In our case, we took content from our main project "Crimea is Ukraine" (available at crimea-is-ukraine.org). We extracted all the materials and converted them into Markdown format — essentially plain text files with structural markup. We wrote specific instructions, and LLM agents powered by large language models began processing them.

In these documents, the language model identifies self-contained concepts or specific individuals. If the model finds self-contained information about a person, concept, or idea in the text, it extracts that entity and writes a concise, encyclopedia-style article, similar to a Wikipedia entry.

After that, processing each source document one by one, articles were generated. Then, the agent cross-references related objects, persons, or ideas by adding internal links, just like Wikipedia, forming a connected knowledge base. Finally, we deployed this as a standalone website at wiki.crimea-is-ukraine.org, which is a subdomain of our main site "Crimea is Ukraine".

Valentyna Troyan: Who else worked on this? When you announced the launch on Facebook, you mentioned a team?

Andrii Bogdanovych: Yes, of course. Initially, six years ago when we launched the project, key contributors included Anatolii Khromov — Head of the State Archival Service of Ukraine (part of the materials came directly from archives); Andrii Shchekun — who was then editor-in-chief of the "Krymska Svitlytsia" newspaper (a large part of our website consists of previously unpublished materials from "Krymska Svitlytsia"); and Serhii Mokreniuk from the NGO "EuroMaidan-Crimea" — who originally authored the project idea. As manager, I brought everything together. Essentially, the four of us built this project.

The Wiki knowledge base itself is built on top of all those materials. Regarding your question about trusting AI output: of course, we must include a disclaimer that AI models can make mistakes and their work must always be verified. The articles generated were strictly based on existing human-written source materials. Once the agent processed all documents and created articles, we ran a second verification pass.

In this second pass, another model (or the same agent with a clean context window without previous memory) re-read the generated articles from scratch, comparing them against the original source materials to verify accuracy and check for hallucinations or out-of-scope information.

Of course, this doesn't guarantee 100% perfection, and any project created with AI requires a conscious approach. However, the probability of errors here is quite low due to these multiple verification iterations.

Valentyna Troyan: Were there any hallucinations?

Andrii Bogdanovych: Of course. During the second pass, when the agent verified the work done by the previous run, errors were caught and corrected. We observed instances of imprecise phrasing or slight misstatements, but overall the output was accurate, and the second pass successfully eliminated a number of errors.

Valentyna Troyan: Interestingly, at what points did it hallucinate or distort information?

Andrii Bogdanovych: I can't point to specific instances right now because there were over 200 source articles initially. Now, on the Wiki project, over 500 articles have been generated based on those sources.

However, any language model fundamentally operates by predicting text — that's its core mechanism. It tries to predict the most relevant next word in a sequence. Invention is inherently part of its architecture. Therefore, completely eliminating hallucinations is impossible by design.

Valentyna Troyan: That's very interesting. The reason I asked about hallucinations is that the topic of Crimea is extremely sensitive, and Russia actively spreads disinformation online. Naturally, when AI trains or processes information, it doesn't automatically distinguish truth from falsehood.

Andrii Bogdanovych: Fundamentally, a model cannot independently distinguish truth from falsehood. It treats the training or provided context data as truth. If a language model is provided inaccurate information, it will generate inaccurate outputs.

My point is that anyone working with AI tools must remain conscious of this and verify the output. In our project, we performed a second verification pass.

Furthermore, every article generated on "Wiki: Crimea is Ukraine" includes a source reference at the bottom linking directly back to our main website. When reading an entry on the Wiki, you are reading a compiled summary of original materials, and you can always follow the link to read and verify the original source.

Valentyna Troyan: Let me remind our listeners that you are listening to Ukrainian Radio. We are joined by Andrii Bogdanovych, founder of digital projects, innovation and AI expert, developer of "Wiki: Crimea is Ukraine". What feedback have you received from your audience and Wiki users?

Andrii Bogdanovych: We launched relatively recently. Feedback has come through personal contacts and social media comments. Overall, people appreciate the approach and welcome this new, engaging resource. The feedback has been positive, with no negative reactions so far.

Valentyna Troyan: Regarding the content, you mentioned unique materials from "Krymska Svitlytsia". What other data was collected for the AI to process?

Andrii Bogdanovych: I was pleasantly surprised myself... Because of the different approach used in this site structure... On the primary "Crimea is Ukraine" site, content is presented as discrete archival articles. On the Wiki, however, one entity extracted by the agent was Albert Einstein — the famous physicist who, as it turns out, conducted research related to Crimea.

I hadn't encountered this fact in widely known public sources. Anyone who hadn't read through every single archival document on our site might never have noticed it. But the LLM Wiki extracted this specific entity and highlighted his connection to Crimea. Such fascinating historical facts emerge naturally when using this structured approach.

Valentyna Troyan: How frequently do you use artificial intelligence in your daily life and in what areas? Where does it help, and where do you remain cautious?

Andrii Bogdanovych: I personally use AI every day for numerous tasks. Regarding caution: one must simply understand that it is a tool. With any tool, you need to understand how it works and treat it accordingly. It is not a magic wand that solves every problem.

You need to leverage its primary strength — the ability to rapidly process large volumes of data — while being aware of its limitation, which is the tendency to occasionally hallucinate or generate non-existent facts. Understanding this allows you to use the technology effectively.

Valentyna Troyan: Has AI ever made a mistake that you initially trusted?

Andrii Bogdanovych: Not blindly, because I make it a rule to verify information. But it is very easy to fall into that trap if you aren't paying attention. AI generates text that sounds remarkably authentic. When an LLM hallucinates, it does so with such plausibility that it can be visually indistinguishable from reality.

Therefore, key facts should always be cross-checked, especially if the AI insists on a specific assertion.

Valentyna Troyan: Do you think "Wiki: Crimea is Ukraine" will gain popularity and active engagement?

Andrii Bogdanovych: I certainly hope so and expect it to bring new life to the project. The main site hadn't been updated in a while, though the collected materials remained valuable. This provides a fresh perspective and interactive approach to that information, uncovering interesting historical threads like the Einstein connection.

Valentyna Troyan: Yesterday you also announced that "Crimea is Ukraine" is now available in Crimean Tatar. That seems like a significant milestone for preserving and promoting the language.

Andrii Bogdanovych: Yes, we launched the Crimean Tatar language version, and we are planning a Turkish version next. For now, that will conclude our language expansions.

Valentyna Troyan: Who contributed to the Crimean Tatar translation?

Andrii Bogdanovych: That was also performed using large language models through multi-pass processing. For translation, we conducted even more iterative passes to correct potential language generation errors.

Naturally, if native speakers spot any remaining inaccuracies, we welcome feedback and will gladly correct them. We have provided an email address on the site for inquiries and submissions.

Valentyna Troyan: Thank you, Andrii. Andrii Bogdanovych was with us on air. Thank you and have a great day!