29&30 October

AI2FUTURE 2025

AI Beyond the Hype - Supported by CroAI

Kraš Auditorium, Zagreb

GET YOUR TICKET

Event Speakers

MEET THE SPEAKERS

View All Speakers

2025 2024

Relja Marković

Managing Director of Digital Transformation at the Addiko Group

Relja Marković is the managing director of Digital Transformation at the Addiko Group level, responsible for all six countries in which Addiko Bank operates. After graduating from the Faculty of Science in Zagreb, he has been working at Addiko Bank since 1999, where during his career he has led the implementation of various digital banking, mobile and web-based solutions and processes. He has been also responsible for overseeing all aspects of defining and implementing digital solutions, from UI/UX processes to architecture, technology and development. He is passionate about researching and integrating new technologies, such as artificial intelligence, and about innovations in improving user experience and business performance.

Davor Runje

VP of Agentic Engineering at Synthpop AI

Davor Runje is the VP of Agentic Engineering at Synthpop AI, where he leads the development of real-time AI systems automating U.S. healthcare operations. A veteran engineer and researcher, Davor is one of the core contributors to Microsoft Research’s AutoGen framework for agentic AI systems and the creator of FastStream, an open-source Python framework for building broker-based architectures with Kafka, RabbitMQ, NATS, and Redis. His work bridges distributed systems, voice technology, and intelligent automation — shaping how AI agents think, speak, and act in real time. Before joining Synthpop AI, Davor co-founded several technology ventures and has been an active advocate for open-source innovation and applied AI engineering. At AI2Future 2025, he will share how agentic voice systems are redefining the intersection of artificial intelligence and healthcare.

Tomislav Strgar

Head of TIS.AI Department

Tomislav Strgar, TIS Group, Head of TIS.AI Department, Project Manager, with 17 years of experience in ICT across various roles. He started as Java developer, continue as technical consultant and IBM DMS & BPM certified specialist working on PoC workshops, systems analysis, architecture and Java development issues resolving, business processes/workflows modeling.

Last several years has focus on business development and project management but without leaving IT and technology overall.

Since 2019 is researching AI field, preparing and managing AI team and AI projects, such as TIS’s AI product in healthcare, SENDD, System for Early Neurological Deviation Detection and Vineyard Angel, AI solution in agriculture/viticulture, and other GenAI projects in different enterprises.

Most recently, Tomislav was honored with the AI Champion Award at AI Summit ADRIA 25, recognizing his outstanding contributions to the regional AI community.

Nikolina Čizmić

IT Business Systems Specialist - Senior at JGL

Nikolina Čizmić, a Master of Engineering, brings experience in diverse IT project implementations. Throughout her career, Nikolina has actively conducted training programs and workshops, integrating them into various project initiatives to ensure knowledge transfer and team empowerment. As a IT Project Manager, she is ledding an AI upskilling initiative in JGL and has completed the Agile Coach certification program, underscoring her expertise in agile methodologies and innovative solutions.

Bruno Ćorić

CTO and Co-founder of Arkensight

Bruno Ćorić is the CTO and Co-founder of Arkensight, specializing in multimodal AI and applied machine reasoning for critical infrastructure. Before Arkensight, he was the lead engineer behind AI systems for solving textual math problems at Photomath. He later worked at Google on the reasoning pillar for Gemini and Google Lens, focusing on advanced perception and context understanding at scale.

Marko Ivić

Portfolio Enabler for Data & AI at Porsche Digital Croatia

Marko is a tech enthusiast with a Master's degree in Computing Science and over 10 years of experience in the field. He has worked as a software engineer, data architect, and technical product owner. Currently, he works at Porsche Digital Croatia, where he is responsible for a portfolio of product teams in the area of Data & AI. In his role, he has the opportunity to contribute to the Data & AI strategy and drive the implementation of some really cool and innovative solutions.

Mihael Španović

Head of Engineering at Protostar Labs

Mihael Španović is Head of Engineering at Protostar Labs, leading ML/AI development across BioTech, industrial vision, and space sector. He specializes in model engineering, pipeline automation, and cloud/on-prem deployment. Mihael previously conducted robotics research at the University of New Mexico and is pursuing an M.Sc. in AI at Johannes Kepler Universität Linz. In his spare time, he’s an avid competitor who participates in ML competitions.

Jakob Triva

Senior Consultant at Solvership

Jakob is a Senior Consultant at Solvership with a strong background in software development and a Master’s degree in Process Computing from FERIT Osijek. His work focuses on creating and managing data pipelines, Kubernetes clusters, and cloud services. With a passion for automation and practical, results-oriented solutions, he strives to make complex systems reliable, efficient, and easy to manage.

Kristijan Lukačin

CEO & Founder of Web Solutions · Co-Founder of Vybrid.ai

Kristijan is a technology visionary, entrepreneur, and AI enthusiast with over 20 years of experience bridging business, design, and software development. He began coding in the early 80s on a ZX Spectrum 48K and has since led enterprise projects in Drupal, marketing automation, and AI-augmented development.

As CEO of Web Solutions, he built one of Croatia’s leading Drupal agencies serving clients worldwide. Through Vybrid.ai, Kristijan pioneers vibe coding — a new approach to product development that blends human creativity with AI tools to drastically shorten time-to-market.

A father of nine, devoted Catholic, and passionate photographer, Kristijan believes that technology should empower people and creativity, not replace them.

Klemen Rizman

Head of the R&D Department at CREAPLUS

Klemen Rizman is the head of the R&D department at CREAPLUS, where they develop custom data and AI solutions for companies. He has been working with machine learning methods and their application to business processes for the past 10 years.

Dora Raštegorac

Senior Associate in the Application Systems Development Department at HNB

I'm Dora, a passionate data scientist with several years of experience. My current degree is a Master's degree in Mathematics and Computer Science, and I am pursuing a PhD with a particular interest in Natural Language Processing.

Bruno Tomrlin

Solution Architect at Infoscope

Solution Architect at Infoscope with a focus on AI solutions. Currently, I’m oriented toward AI projects and the development of intelligent systems, with a background in Documentum, backend development in Java, and the installation, administration, and maintenance of Linux systems.

Check out the Schedule

AGENDA

18:00-20:00

CroAI Meetup x AI2F: Are Self-Flying Planes the Next Big Thing?

Event by CroAI (Croatian AI Association)

Speakers: Ramanathan Srinivasan (Airbus Innovation Center), Dejan Grahovac (Deputy Director of the Croatian Civil Aviation Agency) & Bruno Antulov-Fantulin (Founder & CEO at SKAI LABS)

Location: Planinska ul. 13a (third floor), Zagreb

Event link: https://luma.com/zfmfoe1q

09:30-09:35

Davor Aničić (AI2Future 2025 Programme Committee; CroAI Board Member & Velebit AI CEO)

Opening Remarks

09:35-10:10

Ramanathan Srinivasan (Airbus Innovation Center)

Behind AI Failure

10:10-10:40

Daniel Lampertseder (Alexander Thamm)

Agentic AI - When Artificial Intelligence Acts Independently

10:40-11:10

Enzo Blindow (Prolific)

Beyond the Gold Medal: Steering AI Toward Human-Centric Value and Impact

Centered on the critical gap between models achieving "virtual success" (e.g., winning abstract benchmarks, gold in math olympiads) and delivering real-world, tangible value (economic, utility, and safety) in enterprise settings. I will detail how human data, expert evaluations, and granular feedback loops are tools for moving AI towards alignment with human preferences and institutional goals.

11:10-11:40

COFFEE BREAK

11:40-12:10

Kristijan Lukačin (Web Solutions)

Vibe Coding in Practice: How AI Is Transforming the Way We Build Software

“Vibe coding” represents a new era of software creation — one where AI assists developers, designers, and entrepreneurs to move from concept to production at unprecedented speed.

In this session, Kristijan Lukačin, co-founder of Vybrid.ai (and WEB solutions), demonstrates real-world use cases of AI-driven development, how his team builds production-ready apps in weeks, and what this means for startups and enterprises seeking agility, quality, and innovation.

12:10-12:40

Desislava Sarbinovska (Service Design Digital) & Dimitrina Todorova (Coding Burgas)

AI Needs Designers: Why Cross-Disciplinary Collaboration Is the Missing Link in Responsible AI

As AI systems become more powerful, the gap between technical capability and real-world usability is growing. Too often, AI projects fail — not because the algorithms are wrong, but because the experience is broken, trust is missing, or the system doesn’t serve real human needs.

This talk makes the case for embedding service design from the very start. We'll share how designers work alongside data scientists, engineers, and policy-makers to ensure that AI is not only functional, but also understandable, trusted, and valuable.

We’ll bring this to life with a real-world case study from CodingBurgas, a high school in Bulgaria where 11th-grade students co-created MyPathway — an AI-powered platform for personal and career development. Guided by service design methods, students conducted research, mapped user journeys, and prototyped solutions that serve the real needs of students, teachers, and parents alike. More importantly, they shifted from asking “How do we build this?” to “Why does this matter, and who is it for?”

12:40-13:30

PANEL | AI Regulation

Moderator: Stefan Martinić
Panelists: Andrea Čović Vidović (European Commision), Martina Dragičević (A1 Hrvatska), Marko Đuričić (Visage Technologies)

13:35-14:30

LUNCH

14:30-15:00

Klemen Rizman (Creaplus)

AI agents in Banking: From theory to production

The presentation will clarify the key differences between LLMs, ChatGPT, CustomGPT, and AI Agents, highlighting what truly distinguishes intelligent agents from traditional chatbots. It will also address security and compliance aspects — including data protection (GDPR, banking regulations) and risk management when designing and deploying AI agents, with a focus on what can safely be processed in the cloud versus what should remain on local infrastructure.

Finally, Klemen will present a real-world case study on the implementation of an AI agent system for credit approval processes for individuals — outlining the workflow, achieved results, challenges, and lessons learned.

15:00-15:30

Zlatko Matokanović (Q agency)

AI as a Member of the Dev Team: Collaboration, Control, and Responsibility

Speed without control is dangerous. In this talk, I’ll show how we at Q use AI to accelerate development without creating technical debt and without stifling innovation — through a clearly defined, standardized, and holistic approach to development from idea to delivery.

15:30-16:000

Tomislav Bronzin (Citus)

Tricks, Treats, and Transitions: How Leaders and Employees Perceive AI

AI Assistants/CoPilots offer "treats" to workers – automating repetitive tasks, extracting insights, and increasing individual productivity - while at the same time facing the "tricks" that AI performs, such as bias, hallucinations, and trust issues.

However, a key difference in perception has emerged: workers are intensely experimenting with AI, while many leaders are lagging in the strategic adoption of AI, underestimating its potential and failing to scale pilots into meaningful changes within the company.

Join us in discussing what it takes for leadership to move from cautious observers to fluent practitioners – aligning the application of AI with business goals, investing in skills, and establishing effective governance. All to enable the "transition" that will bridge this gap and fully unleash the potential of the entire organization!

11:40-12:10

Nino Požar (Epam)

EPAM AI/Run CodeMie: Unlocking the Future of AI-Driven Software Development

12:10-12:40

Ena Aničić (Atomic Intelligence)

Unlocking LLM Potential for Text-to-SQL

We explore the design of a BI chat system that translates natural language into SQL, combining enterprise models with smaller open-source alternatives in a modular flow that leverages RAG, tool-calling, and privacy-preserving techniques such as result masking. The discussion focuses on situations when retrieval genuinely improves results versus when straightforward prompting is sufficient, how expert input can refine prompts, and why tailoring models to specific tasks can be more effective than a one-size-fits-all approach. It also considers the cost and performance trade-offs between enterprise and fine-tuned open-source models, while offering a glimpse into ongoing reinforcement learning (GRPO) experiments for SQL generation.

12:40-13:10

Andrzej Marosek (AWS)

Amazon Bedrock AgentCore - Your All-in-One Platform for Enterprise AI Agents

This technical session provides an in-depth exploration of Amazon Bedrock AgentCore's modular architecture for building sophisticated AI agents. We’ll dissect each core module that powers enterprise-grade AI applications: the Memory module for contextual information management, Runtime module for execution orchestration, Gateway for seamless tools integration, Observability for monitoring and debugging, and Identity for security and personalization. Through practical demonstration, we'll showcase how these components work together as we build a fully functional conversational AI application from scratch. Attendees will gain technical insights into how AgentCore's unified architecture eliminates the complexity of agent development while maintaining enterprise-level reliability, scalability, and governance—allowing developers to focus on building differentiating experiences rather than managing infrastructure.

13:10-13:35

Ena Melvan (Metabelly)

Gut Feelings: How AI and Microbiome Science Are Transforming Preventive Health

The gut microbiome is one of the most complex ecosystems in the human body, linked to digestion, immunity, and even mental health. Yet, translating this complexity into actionable healthcare has long been a challenge. In my talk, I will share how advances in AI help us move beyond diversity scores and raw sequencing data toward predictive, clinically useful insights. Drawing on my research and startup experience, I will explore the intersection of AI, microbiome science, and preventive health – and why combining these fields may transform the way we approach healthcare in the years ahead.

13:35-14:30

LUNCH

14:30-15:00

Igor Mijić (Compass AI)

Sizing Up (or Down) for Success: Matching Language Model Size to the Task

We will present an engineering overview of matching the proper models to proper needs, with careful consideration on how an industry with a growing amount of calls to LLMs needs to be able to scale that number in a sustainable way. For example, with Agentic AI on the rise, we see agents being used in most types of automation use cases available, but not all agents need the default, largest and most costly LLM models to run properly. In short, we will present the problem of matching the Language Model to the applications needs, give out some current viewpoints and interpretations from latest scientific/popular/engineering papers and sources, and then showcase examples how appropriate model usage helps both in execution time and in cost.

15:00-15:30

Mihael Španović (Protostar)

Explainable Models, Reliable Outcomes: ML That ships

While machine learning models are rapidly scaling, there are still times when it's advisable to build systems that rely on classic, interpretable models. This talk shares how we used gradient boosting models in three different use-cases: detecting food adulterants, brain data anomaly detection, and satellite anomaly detection. We will cover situations when “simple” beats “state-of-the-art,” the trade-offs we accept, and practical patterns you can reuse for your own projects.

15:30-16:00

Mislav Ćaleta (Hidrocibalea)

From Bottleneck To Innovation: In-House AI for Hydrographic Data Processing Beyond Off-the-Shelf Limitations

Commercial hydrographic data processing software often struggles to keep pace with the evolving needs of modern survey operations. Development cycles can be slow, customization options limited, and performance bottlenecks increasingly evident when dealing with complex and specific needs. To address these challenges, we develop in-house artificial intelligence–driven solutions, specifically tailored to our operational requirements. By designing tools that directly reflect field realities, our in-house AI solutions allow us to streamline workflows, enhance data quality, and deliver results faster.

11:40-12:25

Nebojša Popović (PBZ)

Crash Course - Computer Vision Using OpenCV on Resource Constrained Devices

12:25-13:35

Domagoj Kovač (Sedmi odjel)

Building AI Agents With N8N and Amazon Bedrock

In this hands-on session, we’ll explore how to build simple AI agents by combining n8n and Amazon Bedrock.

We’ll start with an introduction to core n8n concepts - nodes, triggers, and workflows - to understand how automation flows are created and executed. Then, we’ll move on to exploring AI-agent node where we will demonstrate how to integrate Amazon Bedrock models into n8n workflows.

This session is a great resource for developers, solution architects, and business innovators looking to bring generative AI into everyday workflows through low-code automation.

13:35-14:30

LUNCH

14:30-15:30

Jakob Triva (Solvership)

Vibe-Driven DevOps on GCP with Pydantic-AI

How do we build AI systems that don’t just "talk", but reliably act - executing complex operations safely and predictably?

This technical session explores the foundations of autonomous AI agents through the lens of MCP (Model-Context-Protocol) and Multi-Agent architectures, where “Vibe-Driven DevOps” is the powerful outcome.

Through a live demo on Google Cloud Platform, Jakob will showcase these concepts in action as he sets up an entire infrastructure environment - simply by conversing with an LLM-powered agent.

Join this session to learn how to architect the next generation of AI-driven tools, built for precision, safety, and scalability.

10:00-10:30

Hrvoje Odak & Matej Lončarić (RTL Hrvatska)

AI at RTL: From Early Steps to Real Impact

RTL is in the early stages of our AI and data transformation, driven by strong leadership from CME and positive momentum within our organization. Our strategy is to focus on practical tools that enhance business processes and can be adopted quickly, and we’ve successfully deployed some of them across the organization, from Newsroom, Content acquisition and planning, digital marketing, CRM and Customer service…

At the same time, we run targeted pilot projects and internal experiments, and we prioritize education, offering foundational AI training to colleagues to foster readiness and engagement. While our maturity is still developing, executive alignment, effective tools, and internal enthusiasm are already generating real impact. At AI2Future, we will share how this coordinated, people-first approach is shaping our transformation and helping us move beyond hype towards tangible business results. Our assumption is that there are a lot of regional companies that are in the same situation as we are, so let’s share ideas and experiences and learn from each other.

10:30-11:00

Tomislav Strgar (TIS Grupa) & Nikolina Čizmić (JGL)

The Myth of "SIMPLE" GenAI Implementation...Real Use Cases Uncovered

11:00-11:30

COFFEE BREAK

11:30-12:00

Sandro Stojaković & Tomislav Herceg (Sedmi odjel)

Beyond Buzzwords: How AI Is Actually Helping Our Clients

In this session, we'll move past the hype and share our case studies. Sandro will introduce Sedmi odjel, while Tomislav will walk you through three case studies - showing where AI created value.

12:00-12:45

PANEL | GDPR

Moderator: Marija Bošković Batarelo (Parser compliance)
Panelists: Martin Morava (Devot), Karlo Paljug Zorko (AZOP), Ivan Gazić (Kent banka)

12:45-13:10

Ivo Lukač (Netgen & Konfica)

Mingle AI B2B matchmaking: find qualified business connections with AI

Conferences and other business events can provide hundreds or thousands of potential contacts, but attendees can only have time for a few meaningful conversations. Too often, valuable opportunities are missed due to time limits, unclear fit, and lack of structure in evaluating interactions. We are building MingleAI, an AI-powered system that simulates networking conversations between attendees, evaluates the potential for each, and recommends qualified, high-value matches, increasing ROI of participating in events.

13:10-13:30

Bruno Tomrlin (Infoscope) & Dora Raštegorac (Croatian National Bank)

How GenAI Is Changing PR Media Analysis

13:30-14:30

LUNCH

14:30-15:20

PANEL | AI in Organizations: What really works?

Moderator: Andreja Dobra (A1 Hrvatska)
Panelists: Ante Laušić (Solvership), Duje Bralić (Valamar), Relja Marković (Addiko banka), Marko Katavić (Moneybox)

15:20-16:05

Startup Block

Dario Deković (Ultrax) - Building the AI Assistant Coach: An Agentic Approach to Sports Analytics

Bruno Ćorić (Arkensight) - Enabling Critical Infrastructure Resilience with Multimodal AI

Filip Kuzminski (DeltaSort) - Can We Make the Dirty Clean Again?

Building the AI Assistant Coach: An Agentic Approach to Sports Analytics

Learn how we built an AI Assistant Coach to automatically generate sports analytics reports while tackling critical LLM pitfalls. This session provides a practical blueprint for creating a factually-grounded AI system capable of handling complex, real-world data.

I'll walk through our architectural solutions for three common LLM challenges we encountered. To solve poor numerical reasoning, we bypassed the model's limitations by building a suite of deterministic analysis tools. To overcome context window limits that caused players to be missed in large datasets, we implemented an Orchestrator-Worker pattern where individual agents analyze each player in parallel. Finally, to ensure factual accuracy and report consistency, we built a layer that calculates predefined analysis provided by real coaches. These analysis ensure our reports strike a balance between exploration and consistency.

Enabling Critical Infrastructure Resilience with Multimodal AI

This session explores how Arkensight applies multimodal AI to transform visual data into operational intelligence for critical infrastructure. From ports and railways to energy and utilities, we’ll showcase real-world deployments that reduce downtime, enhance safety, and drive autonomous asset management.

Can We Make the Dirty Clean Again?

Every product we use begins as material extracted from the Earth — yet, once discarded, it’s treated as worthless. This talk explores how robotics and automation are closing that loop, transforming waste back into valuable resources. Through innovation in AI-driven sorting and accessible automation, we can rethink how we handle the materials we once took for granted — and ask the crucial question: can technology help us make the dirty clean again?

09:30-10:00

Davor Runje (Synthpop AI)

From Speech to Action: Building Real-Time Voice Agents to Automate Healthcare

Voice is becoming the most natural interface between humans and machines — and in healthcare, it can be the most transformative.

In this talk, Davor Runje, VP of Agentic Engineering at Synthpop AI, unveils how his team built a proprietary real-time voice agent engine that automates complex administrative workflows across the U.S. healthcare system. Drawing on the latest advancements from OpenAI’s and Google’s real-time voice models, as well as principles from agentic AI frameworks like AutoGen, the system transforms spoken interactions into intelligent, compliant, and fully automated actions.

Davor will share lessons learned from building low-latency, secure, and reliable voice agents that work in production environments - from architecture to deployment - and how this technology is reshaping the “business of care.”

The talk bridges engineering depth, AI design, and real-world impact, showing how speech is becoming the new API for healthcare.

10:00-10:30

Borna Bejuk (A1 Hrvatska)

Who you gonna call? Nikša!

In today's fast-paced world, we all want a helpful response to our problem as soon as possible. That's why we've developed the Nikša voicebot, an AI companion that is always available and able to engage in natural language conversations. Nikša is trained to assist with the most common customer inquiries, such as activating tariffs and checking the current balance. Join us to learn how our Voicebot ecosystem operates and what are the components that make it a valuable resource for our customers.

10:30-11:00

Marijan Butkovic & Bruno Brozović (Span)

From RAG to Agentic RAG: Framework Selection for Enterprise AI Applications

Traditional RAG is great for retrieving documents, but what if your users need to analyze Excel files, query databases, or work with data from multiple sources at once? This is exactly what our enterprise clients ask for - they want these capabilities and they choose our solution because they trust us with their data privacy.

This talk is about Agentic RAG - basically AI systems that can use multiple tools on their own. We'll share what we learned from testing different frameworks with actual proof-of-concepts. We compared how well they handle tools, and combining information from various sources.

You'll learn what to look for when choosing a framework, common implementation patterns, and the real trade-offs we discovered.

11:00-11:30

COFFEE BREAK

11:30-12:00

Matej Kuhar (Aircash)

Baltazar: Architecting Trust

On portals and in the news, we increasingly read about various financial frauds, from so-called romantic and investment scams to social engineering. With the development of digital technologies, users can execute transactions much faster, which also increases the risk of misuse and makes detailed transaction checks more difficult to perform in real-time. We therefore developed significant upgrades to Aircash's in-house transaction monitoring system to further protect our users. The architecture of the system is event-driven, where every executed transaction is checked in real time. The system calculates numerous parameters about the transaction itself and about the user’s transaction history and provides an assessment of whether the transaction is potentially malicious. This architecture also enables the integration of advanced system components, such as global traffic analyses and AI models.

12:00-12:30

Disha Panchal & Valentin Cojocaru (Datatonic)

How to Build, Monitor, and Govern Production-Ready AI Agents

The journey from use case experimentation to a reliable, scalable, and compliant Agentic AI system presents a new set of challenges. How do you move beyond the "happy path" to build for the complexities of the real world? And how do we ensure these complex, autonomous systems operate reliably, safely, and in full compliance with enterprise governance policies?

We'll explore how to design a scalable AI platform that provides a collated view of performance, cost, and risk, with tailored governance views for both central oversight teams and individual use-case owners.

12:30-13:00

Daria Štefić (Solvership)

LLMs Evaluation: The New Paradigm

Just a few years ago, evaluating AI systems was all about familiar metrics like accuracy, precision, recall, and false positives. With the rise of large language models (LLMs), evaluation has become far more complex—both in academic research and even more so in business applications.

In this talk, we’ll explore why LLM evaluation is so challenging, highlight common pitfalls, and share practical solutions you can apply today. We’ll demonstrate how ready-to-use Google Cloud tools can help assess key aspects of LLMs, and compare different evaluation methodologies, showing when and how each is most effective.

13:00-13:30

Maria Katić & Katarina Kovačević (Pontis Technology)

Computer Vision Applications in Pharma, 3D Reconstruction and Geospatial Mapping

This presentation will focus on three practical applications of computer vision across different domains – from pharmaceutical quality control to geospatial navigation.

These are the use cases our CV team has worked on in the past year working as Machine Learning Engineers.

  • Pharmaceutical Industry: Object Detection and Classification
  • Detecting imperfections in vials and ampoules to ensure quality in production lines
  • 3D object reconstruction: Camera pose estimation and NeRF (Neural Radiance fields) technology
  • Using camera pose estimation and neural radiance fields (NeRFs) to generate detailed 3D models of small objects
  • Geospatial Mapping: Feature extraction, feature matching and SLAM
  • Leveraging drone footage to estimate GPS coordinates and support navigation.
13:30-14:30

LUNCH

14:30-15:00

Nedim Muhamedagić (Bloomteq)

Agility over Complexity: AI Agents that Transform Supply Chains

What if your competitive edge didn’t require building a giant AI ecosystem? In this keynote, we uncover how simple, cost-effective agentic flows and focused RAG modules can plug into existing supply chain platforms, unlocking efficiency, resilience, and innovation with a fraction of the investment.

15:00-15:30

Zlatan Ajanović (RWTH Aachen University)

Machine Learning and Reasoning for Intelligent Robot

Robots are getting better at seeing, grasping, and moving, but still struggle to understand what we really want them to do. A simple request like “bring me the milk” requires connecting language, perception, and physical action across many levels of reasoning. The robot must understand the goal, plan a sequence of steps, and adapt its motion safely in a changing world.

In this talk, I will explore how recent progress in machine learning and reasoning brings us closer to truly intelligent robots. By combining large language and vision–language models with new planning and learning methods, we can teach robots to interpret human instructions, reason about complex tasks, and learn from experience. These advances allow robots to move beyond rigid programming toward flexible, open-world behavior, learning to use language as a bridge between human intent and physical action.

15:30-16:00

Marko Ivić (Porsche Digital)

The Scalability Imperative: Achieving Scalability in Architecture and Organisation

Creating a scalable structure is a prerequisite for the efficient application of AI. This session will break down the complexity of the automotive data landscape through examples and demonstrate why organisational scalability must be your primary focus. We will deep-dive into both technical and strategic components - including data products, self-service platforms, governance, and decentralisation to provide you with a comprehensive understanding of the strategic requirements for scaling enterprise AI.

10:00-10:30

Chris Thomas (Alan Turing Institue)

Building Capability for AI Regulation and Standards

This presentation will cover the importance of capability building for effective AI governance, addressing capability building strategies for AI regulation and AI standards development and implementation. The presentation will focus on the recently developed AI Regulatory Capability Framework and work by the AI Standards Hub.

10:30-11:00

Nikola Ljubešić (Štefan Jožef, Ljubljana)

Natural Language Understanding Is a "Solved" Task, Right? On the Long Tail of Language Variation

The recent wave of advances in artificial intelligence has started with language modelling, and has since then moved on to problems of reasoning, world models, and other tasks beyond language understanding. Large language models have redefined what machines can do with text, leading to a widespread perception that natural language understanding is now a "solved" problem. Yet, language is vastly more diverse than the narrow, standardized domains on which these models are typically trained and evaluated on. True mastery of language requires handling the long tail of linguistic and communicative variation -- the countless dialects, registers, and complex situations that characterize real human interaction.

In this talk, I argue that current technologies, despite their apparent fluency, fail to capture this complexity. I will illustrate the problem through a series of recent experiments on South Slavic languages and dialects, across both textual and spoken modalities. These include the DIALECT-COPA benchmark for dialectal natural language understanding, the Mići princ dataset for speech recognition in the Chakavian dialect, and the still-developed Global-PIQA dataset for physical commonsense reasoning across standard and non-standard varieties. In all cases, state-of-the-art LLMs and speech models show sharp performance drops when faced with dialectal or less-represented forms of language -- indicating that their apparent "understanding" remains superficial and unevenly distributed.

These findings suggest that what has been "solved" is only a small, standardized slice of human language. Addressing the long tail of linguistic diversity is not just a matter of fairness or inclusivity, but a prerequisite for building systems capable of real-world applications outside the safe havens of benchmarks. And this journey has just started.

11:00-11:30

COFFEE BREAK

11:30-12:00

Josip Šarić (Faculty of Computer and Information Science, University of Ljubljana)

Born To Be Wild or Farmed? Fish Origin Identification Using Deep Learning

Fish origin identification is critical for aquaculture management, particularly for assessing ecological and genetic impacts when farmed fish enter wild populations. It is also essential for ensuring transparency and preventing fraud in fish markets. Geometric morphometrics provides a simple and cost-effective approach to origin identification through shape analysis. However, manual digitisation of landmarks remains prone to variation and introduces measurement error that can obscure or distort the biological signal. This study addresses these challenges by proposing a novel deep learning method for automatic landmark placement in gilthead seabream, a key species in Mediterranean aquaculture. Our automated system detects landmarks with sub-millimetre accuracy and correctly determines the fish origin in 95% of test cases.

12:00-12:30

Paul Duncan (National Physical Laboratory - NLP)

Metrology for Trustworthy AI: Measuring What Matters

AI is transforming sectors such as healthcare, finance, environmental monitoring, manufacturing and transportation by advancing data analysis, predictive modelling and automation. As these technologies become embedded in everyday operations and decision-making, confidence in the responsible and intelligent use of data and AI becomes increasingly critical. This presentation explores how metrology, the science of measurement, and the work of National Metrology Institutes can support the development of trustworthy AI systems, foster public confidence and contribute to the creation of robust standards for AI adoption.

12:30-13:00

Marko Čuljak (FER, Zagreb University)

A Bag of Tricks for Effective Text Embeddings with Large Language Models

When tackling high-value tasks such as information retrieval, text clustering, and sequence labeling, modern natural language processing (NLP) systems rely on dense low-dimensional text embeddings. Owing to the attention mechanism and the masked language modeling training objective, encoder-only language models (LMs) such as BERT and RoBERTa are well-suited for producing high-quality text embeddings and have long been topping the leaderboards. On the other hand, decoder-only LMs such as Gemini, GPT, and Llama excel in text generation but generally produce weaker text embeddings than encoder-only models with a similar number of parameteres. However, since many real-world tasks can be successfully modeled as text-generation tasks, decoder-only LMs have been scaled to much higher parameter counts than encoder-only LMs, equipping them with more processing power. This shift in focus has prompted the broader NLP community to devise approaches that effectively and efficiently utilize the power of decoder-only LMs for producing high-quality text embeddings. This talk will provide an overview of these approaches and their technical details, and present the results of a recent study on the topic, which is currently under peer review.

13:00-13:30

Blum Institute, Sarajevo

13:30-14:30

LUNCH

14:30-15:00

Damir Korenčić (Ruđer Bošković Institute)

Multi-Agent Systems for Machine Translation

Large Language Models and multi-agent systems have significantly advanced machine translation (MT) capabilities in recent years. This talk examines several modern multi-agent MT systems, including two that I helped develop, and provides an overview of recent WMT translation challenges. The presentation will also connect MT-specific observations to general concerns that arise in LLM-based systems.

The power of community

PARTNERS 2025

Top

Platinum

Gold

Community Partners

Interested in partnering with AI2Future?

Relevant info

About AI2Future

Venue

Kraš Auditorium
Ravnice 48, Zagreb

DATE & TIME

29-30 October
9:00 - 17:00h

"AI2Future is conference made by community for community. Since 2017, we are going beyond the hype and aiming to cover practical, real-world use cases interesting for engineers and researchers as well as practitioners from other industries. This year’s lineup of speakers will be the strongest so far, so make sure not to miss it!"

- Marko Velić (Software Engineer at Google)

"AI2Future isn't just a conference; it's where Croatia's AI vision takes shape. By bringing together global experts and local innovators, we're setting the stage for a future where AI empowers every sector of our society."

- Lorena Barić (Executive Director - CroAI)

"AI2Future is a great place to learn and connect with experts in the field. The conference provides just right mix of deep technical talks, visionary keynotes, and honest discussions about trends and challenges. Can wait for the next one!"

- Domagoj Kovač (Business Development Manager - Sedmi odjel)

Behind the scenes

Program Committee

Davor Aničić

CEO and Co-founder @ Velebit AI, Board Member @ CroAI

Seasoned business leader with a product management background leading a tiny, efficient and very experienced team of Velebit AI that implements B2B solutions and delivers AI services for online marketplaces, publishing, e-commerce, gaming, biotech and many other industries. Serial entrepreneur starting in the early internet days but also spent five years in academia and 15 years in the corporate environment in various roles across industries like telecommunications, online publishing, and technology transfer. Driven to understand how AI will interact with humans and change our livelihoods in the years to come.

Davor Runje

Head of OSS Engineering @ AG2

Davor is the Co-founder/Chief Scientist at AIRT, an AI Technology company that helps finance companies to use big data techniques on (pre)process transactional data and dramatically improve accuracy of all predictive models created and used by the internal data science team.

Before co-founding DRAP in 2008, he co-founded PlayMedia Systems in 1997 and launched AMP MP3 technology on the global market, and helped launch BabyWatch (now called Bellabeat) in 2013, a digital health wearable company. He also teaches machine and deep learning at the HUB385 Innovation center.

As a PhD intern working with Yuri Gurevich at Microsoft Research, he worked on applications of the interactive ASM thesis on multiprocessor/multicore programing. He designed, implemented and transferred technology to a product group of a system for execution of structured concurrency, later named the Task Parallel Library of .NET framework. Received the SSCLI and Phoenix 2005 award by Microsoft Research as one of the best 16 research projects in international competition.

Inventor and lead developer of several DRM based and DRM free systems for digital audio distribution based of AMP® MP3 decoding engine for clients including DMX Music, STMicroelectronics and Napster.

Matko Bošnjak

Senior Research Scientist @ Google DeepMind

Marko Velić

Software Engineer at Google

Marko is an Software Engineer at Google. Previously he worked at Photomath, Meta (Facebook), Styria, LEGO (as consultant), University of Zagreb and has entrepreneurship experience. He won many awards including the ones from Microsoft and NVIDIA. Marko is also lecturer at Algebra University. He is interested in ML research and practical applications in real life scenarios.

Jan Šnajder

Full professor @ FER


ABOUT THE ORGANIZER

Experienced at event management, ideation and implementation, Eventa 01 is a trusted partner to a wide range of Tech, IT, Data & Analytics and AI organizations and individuals.

Aiming at creating forward-looking and productive environment, we build communities through events, project collaborations, matchmaking and business development.

SUPPORTED BY CroAI

CroAI-building a bold AI future through fostering community, collaboration, and innovation.

See you in October!

GET YOUR TICKET

Blind Bird

140EUR+ VAT

*LIMITED NUMBER TO 02.05.

Early Bird

210EUR+ VAT

*FROM 02.05. TO 13.10.

Let’s Keep in Touch

We are sending your data!

Thanks for subscribing!

Your data is sent.

Ooooops!

Something is wrong. Please get in touch via e-mail: info@eventa01.com.