Aleksandar Tomašević, Researcher, University of Novi Sad

This talk presents a technical framework for simulating online social interactions using agentic AI and uncensored large language models to rigorously evaluate community moderation strategies. We use Dolphin 3.0 (based on Llama 3.1 8B) to generate realistic, unfiltered discourse, integrating it into Y Social simulation platform where AI-driven agents interact autonomously. These agents, modeled with diverse personality traits, ideological biases, and toxicity levels, create a dynamic and challenging environment. To facilitate the testing of moderation interventions, we employ ToxiGen-RoBERTa for fine-grained offensive content detection and OpenAI’s Omni-moderation model for broader unwanted content flagging. This allows for real-time intervention within the simulation, mirroring human moderation practices and enabling the observation of their effects on emergent community behaviors. We demonstrate how this framework provides a powerful tool for developing and refining nuanced moderation strategies.

Andrija Sagić, Head of Digital Dev., “Milutin Bojic” Library / Europeana

In recent times we witness that open weight models are getting smaller and competing with much larger, also its multilingual support is widening which allow us to use it in our own language. This presentation will explain a process to build your own AI Agent that can be used on low GPU resource machines. Explaining a process the main technical issues raise when implementing a local language support, models and tokenizers. How to select appropriate components to build an AI Agent regarding the machine capacity and how to plan the use regarding to specific needs.

Salavat Garifullin, ML Researcher, OpenDataScience

[english] InterSlavic was designed as an interlingua understood by speakers across all Slavic countries, regardless of their native language. In this talk, I introduce novel Generative AI language models specifically fine-tuned to produce and understand InterSlavic. These models assist users in translating to InterSlavic and facilitate learning the language. Presentation slides will be provided in both English and InterSlavic. If possible, I would prefer to deliver the talk in InterSlavic to directly demonstrate the language’s practical effectiveness.

[interslavic] Medžuslovjansky jezyk byl stvorjeny kako interlingua, ktoru razumějut govoriteli iz vsih slovjanskyh krajin, nezavisno od jih rodnogo jezyka. V toj prezentaciji predstavljajemo nove generativne jezyčni AI modely, speciaľno obučene za tvorjeńje i razumeńje medžuslovjanskogo jezyka. Te modeli pomagajut koristnikam prevoditi na medžuslovjansky i oblegčajut izučeńje jezyka. Slajdy prezentacije budu dostupne i na angliskom, i na medžuslovjanskom. Ako bude možno, ja byh htěl provesti prezentaciju na medžuslovjanskom, daby neposrědno pokazati praktičnu efektivnosť jezyka.

Nino Požar, Data Scientist, EPAM

This talk explores how a RAG model can be developed from a proof of concept (PoC) to a fully functional production application. The focus is on key challenges and best practices in implementing RAG systems in business applications, including evaluation methods and techniques for improving generative responses. Special emphasis is placed on advanced approaches such as hybrid search, reranking, and fine-tuning to enhance the relevance and accuracy of answers. The goal of the presentation is to cover real-world project experiences, implementation challenges, and strategies for successfully deploying RAG solutions in a production environment.

Dejan Đekić, Owner, Functor

Data Maturity is key to successful AI adoption. If the data quality in your systems is of subpar quality, no amount of AI can help you enhance your business.
Quality Data Infrastructure is the foundation of your AI initiatives.

The presentation will give you answers to the questions of what the State Data Center in Kragujevac is as a key infrastructure for the digitalization process, why in Kragujevac, who are the users of the State Data Center, what are the levels of security and reliability, the certification process for the highest class 4 according to the EN50600 standard, what is the National Artificial Intelligence Platform, who are the users of the platform, platform capacities and expansion plans.

Borko Mihajlović, Senior Frontend Developer, Material Exchange

Discover how Frank, an autonomous AI agent system, is transforming the fashion industry’s material sourcing process. By combining advanced document analysis, brand context understanding, and intelligent material matching, Frank bridges the gap between trend books and material selection. Our system processes complex trend books using computer vision and natural language processing, understands brand DNA through historical data analysis, and autonomously manages sourcing projects.

Nenad Resimić, Test Lead, Levi9

Discover how AI can revolutionize test automation workflow and dramatically reduce development time. This presentation demonstrates how Cursor AI can accelerate Playwright test automation, turning hours of coding into minutes of efficient implementation. Through practical examples, learn how AI assistance can help generate test scripts faster, handle complex scenarios more efficiently, and streamline testing process. Join me to explore how AI tools can transform approach to test automation, making it faster and more productive than ever before.

Marina Mijatović, CEO, AngryLynx

In this talk, we present a case study of applying generative AI to bridge the gap between advanced clustering algorithms and actionable business insights. Working with a dataset of 300,000 user profiles and over 300 features, we used the KPrototypes algorithm to uncover behavioral customer segments across mixed data types. We then integrated generative AI models to transform raw statistical summaries of each cluster into clear, human-readable personas tailored for business stakeholders.