From a Destroyed City to a Research Hub: The Story Behind Stabilarity

From a Destroyed City to a Research Hub:
How Stabilarity Came to Be

Oleh Ivchenko & Dmytro Grybeniuk  ·  March 2026  ·  Future of AI Series

The Students #

The story starts in a classroom, as most research stories do — though this particular classroom was unofficial. Around 2019, Oleh Ivchenko began running supplementary IT courses at Odessa National Polytechnic University. Not because the institution asked him to, but because the gap between what students were being taught and what the industry actually needed had become too large to ignore. He recruited three other lecturers, organised practical modules in Web development and Machine Learning, and eventually built these into a structured learning track under ASTIN University, which he still leads as head of the ML and Web practicals.

More than one hundred students came through those courses. Some went into enterprise software. Some into startups. Some stayed in research. Dmytro Grybeniuk was one of them.

Odessa, a Bot, and a $10,000 Grant #

In 2021, Ukraine’s Ministry of Digital Transformation co-organised Hack Locals 2.0 with the United Nations Development Programme. The brief: design digital tools that strengthen security and civic services in Ukrainian territorial communities. Oleh and his collaborators entered with OTG-bot — a Telegram-based service router that let residents submit requests and had them automatically dispatched to the appropriate municipal department. The channels included administrative services, emergency contacts, and one module that would turn out to matter more than anyone expected: missing and absent persons lookup.

The hackathon jury planned to award three grants. They awarded four. OTG-bot was the additional winner, receiving a $10,000 USD UNDP grant (UNDP Ukraine, 2021^[2]). The jury’s evaluation noted the civic security value of automated routing to services that most residents did not know existed or could not reliably reach.

The pilot city was Popasna, Luhansk Oblast. In spring 2022, Popasna was entirely destroyed. The municipal infrastructure OTG-bot connected to ceased to exist along with most of the city. The codebase survived. The missing-person intake module — structured, searchable, routing-capable — survived as an idea in search of better infrastructure. Four years later, it became the ScanLab Tattoo Identification system described in our 2024 research^[3]. But we are getting ahead of ourselves.

The Algorithm That Predicted Viral Content #

While the missing-person thread was going quiet, another thread was accelerating. Dmytro Grybeniuk had been working on a research question that sounds simple and turns out not to be: can the emergence of a cultural signal — a sound, a video format, a creator — be modelled as a predictable phenomenon before it goes viral?

His answer, developed through academic work including a 2024 paper on forecasting the behaviour of social objects and a presentation at the XII International Scientific Conference in Tokyo, was: yes — but only with a specific hybrid architecture. Not black-box deep learning. Not pure statistics. A combination of time-series decomposition, correlation-guided feature selection (his team found that likes strongly predict follower growth; shares do not), and lightweight ensemble models — Random Forest and XGBoost — that remain accurate under the volatile conditions of a platform where trends disappear as fast as they emerge.

That research became Flai Analytics (flaidata.com^[4]), a Ukrainian-Estonian startup co-founded by Dmytro focused on AI-driven TikTok prediction. In November 2023, Flai was selected as an Alpha startup at Web Summit Lisbon — one of the most competitive startup showcases in the world. At booth A834, Dmytro’s team demonstrated a live prediction dashboard. A TikTok video was shown alongside a model output: 70%+ probability of reaching 1 million views within 48 hours. Two days later it hit 1.3 million. Technology.org covered it as evidence that the “chaotic” dynamics of short-form video could be modelled. TechBullion described Dmytro as someone who had built “a crystal ball for social media” — a description he is somewhat uncomfortable with, because the actual mechanism is a precisely specified statistical architecture, not prophecy.

The distinction matters to him because the science is generalisable. The same models that predict viral sounds have been applied to forecast the adoption of new pharmaceutical products — explored in Tokyo — and to other domains where social systems exhibit emergent behaviour. TikTok was the laboratory. The research questions were always larger.

Gromus: When a Producer Called #

The work did not stay abstract for long. Yuriy Nikitin — one of Ukraine’s most prominent music producers — was looking for a way to systematise something that the music industry had always treated as instinct: knowing which sound would break through, and when. He met Dmytro and the ONPU team, who had built neural network models capable of analysing TikTok big data at a scale no existing tool could match. The partnership produced Gromus — from “growing music,” or alternatively “nurturing music,” depending on which translation you prefer (Mixmag Ukraine^[5]).

Gromus (gromus.ai^[6]) offers AI-driven recommendations, hashtag targeting, playlist curation, and analytics for music creators across TikTok and Spotify. The engineering is Dmytro’s; the industry relationships are Nikitin’s. The model that predicts which sounds will gain traction is a direct descendant of the academic forecasting work. This is what applied AI research looks like when it is done seriously: a theoretical framework, a real-world test environment, a commercial product, and a research publication — all in coherent sequence.

The Research Community Forms #

By 2023, the network around Oleh and Dmytro had grown into something more structured. The ASTIN University courses were running continuously. Research collaborations had extended to include Iryna Ivchenko on intellectual data analysis and data mining taxonomy. Publications were accumulating — economic cybernetics, machine learning for pharma portfolio optimisation, AI economics, cost-effective enterprise AI, anticipatory intelligence. The work was real and specific: each paper addressed a defined problem with a defined methodology, academic, registered with DOIs, cited by others.

What was missing was a home. Academic work published across different journals, platforms, and preprint servers is harder to find, harder to cite, and harder to build on than work that lives in a coherent, accessible place. The research community had the content. It needed the infrastructure.

Stabilarity Research Hub (hub.stabilarity.com) was built to be that infrastructure. Not a personal blog. Not a vanity site. A structured repository of original research — each article academic internally, Zenodo-registered with a DOI, category-organised by research series, and published under the authors’ real names with full attribution. The Medical ML Diagnosis series (Oleh Ivchenko, 35 articles and counting) documents ML-assisted diagnostics in clinical settings. The Anticipatory Intelligence series (Dmytro Grybeniuk and Oleh Ivchenko) maps the gaps in the academic literature on anticipatory AI systems. The Holistic Portfolio Framework series formalises the DRI/DRL methodology developed for Oleh’s PhD dissertation in Economic Cybernetics at ONPU. The tools — ScanLab, the AI Use Case Classifier, the ROI and Risk calculators, the Adoption Gap Monitor — are not marketing materials. They are implemented research outputs, each backed by a specification and a publication.

The API was added when it became clear that the tools were being used by researchers who wanted to integrate them into their own workflows. Free for all community members, with no commercial intent. The decision to open it was not difficult: the whole point was accessibility.

The Tattoo Module — and Why It Exists #

Which brings us back to Popasna.

In 2024, while building ScanLab’s diagnostic infrastructure, we returned to the missing-person module from OTG-bot. The original intake form had a free-text field for identifying features — tattoos, scars, distinguishing marks. Nobody had automated it in 2021 because the technology to do so reliably was not accessible at that scale. By 2024, it was.

The ScanLab Tattoo Identification system extracts feature vectors from tattoo photographs — dominant colours, edge density, texture complexity, style classification — and matches them against a voluntary GDPR-compliant registry. It does not require a sketch artist. It returns a quantitative similarity score. It exposes a REST API. Images are not retained after analysis. The design was completed internally and has been operating since late 2024.

In February 2026, a government institution launched a national-scale tattoo identification service for the same purpose — locating missing persons in the context of armed conflict, with over 90,000 registered cases. The service uses manual sketch submission and text-based filtering. Ours uses photograph upload and automated feature extraction. The two approaches independently arrived at the same core schema: sequential sketch numbers, category/body-location taxonomy, linked records per individual. That convergence confirms the design is sound. The technical difference confirms there is still work to do.

We hope to renew OTG-bot. The communities that the original pilot served — in forms that no longer exist as they did, in places that no longer look as they did — are precisely the communities that need automated, low-infrastructure identification tooling most. The research is done. The implementation is live. The remaining question is deployment.

What This Hub Is #

Stabilarity is not a company. It is not a brand. It is the accumulated output of a specific group of people who decided to do the research seriously and make it public.

Oleh Ivchenko (ORCID: 0000-0002-9540-1637) is a PhD candidate in Economic Cybernetics at ONPU, Innovation Tech Lead at a major technology consultancy, author of the HPF-P framework for pharma portfolio optimisation, Scopus-indexed researcher, and the person who has taught more than a hundred students how to build things that work and document things that matter. Dmytro Grybeniuk (ORCID: 0009-0005-3571-6716) is the researcher behind Flai Analytics and Gromus, a co-author on the Anticipatory Intelligence series, and one of the few people who can credibly describe how to forecast viral social dynamics from first principles — because he built the system that does it, validated it at Web Summit Lisbon in front of a thousand people, and published the methodology. Together they have co-authored Scopus and Springer-indexed papers, presented at international conferences, and built tools that real researchers use.

The hub exists because research should be accessible, reproducible, and useful — not locked in institutional repositories, not behind paywalls, not buried in PDFs that nobody finds. Every article here has a DOI. Every tool here has a specification. Every API call here is free.

That is the whole story. It started with students in Odessa, a Telegram bot in a city that no longer exists, and a question about what happens to a missing-person record when the city service it routes to is destroyed. It has not finished yet.

Referenced Works and Resources #

• UNDP Ukraine (2021). Hack Locals 2.0: Digital Transformation Ministry and UNDP Choose Hackathon Winners. undp.org/ukraine^[2]
• Grybeniuk, D. (2024). Forecasting Behavior of Social Objects. Khmelnytskyi National University / XII International Scientific Conference, Tokyo.
• TechBullion (2025). How a Ukrainian AI Researcher Built a “Crystal Ball” for Social Media. techbullion.com^[7]
• Technology.org (2025). How a Ukrainian-Estonian Startup Is Predicting the Future of TikTok. technology.org
• Mixmag Ukraine. Gromus — Growing Music: Yuriy Nikitin and the ONPU Team. mixmagukraine.com^[5]
• Flai Analytics. flaidata.com^[4]LinkedIn: Booth A834, Web Summit Lisbon, November 2023.
• Gromus AI. gromus.ai^[6]
• ASTIN University. university.astin.co^[8]
• Ivchenko, O. ORCID: 0000-0002-9540-1637
• Grybeniuk, D. ORCID: 0009-0005-3571-6716
• Ivchenko, O. (2026). Tattoo-Based Emergency Patient Identification: From Internal Research to Public Deployment. DOI: 10.5281/zenodo.18929669

1. Stabilarity Research Hub. (2026). From a Destroyed City to a Research Hub: The Story Behind Stabilarity. doi.org. dtir
2. Rate limited or blocked (403). undp.org. tt
3. Stabilarity Research Hub. Tattoo-Based Emergency Patient Identification: From Internal Research to Public Deployment. tb
4. flaidata.com. flaidata.com. v
5. Знайомтесь, Gromus: впливовий продюсер Юрій Нікітін створив інноваційний сервіс для просування артистів – ТРЕНДИ – Mixmag Ukraine. mixmagukraine.com. v
6. Gromus.AI. gromus.ai. l
7. How a Ukrainian AI Researcher Built a “Crystal Ball” for Social Media – And Why the World Is Watching – TechBullion. techbullion.com. v
8. university.astin.co. university.astin.co. v