The 2025 GAIA Conference

Omics is an interdisciplinary field examining the interactions and functions of biological components across multiple layers, such as the genome, transcriptome, proteome, and metabolome. By integrating multimodal AI approaches, we can derive critical insights from these diverse areas, uncovering intricate relationships among genes, RNA, proteins, and metabolites in living organisms. Our exploration into developing multimodal strategies for omics starts with Transcriptomics, the most extensively studied domain, where we look at gene expression to see which genes are active and their level of activity.

In this presentation, Balaji Selvaraj will guide us through the transformative shift from traditional bulk sequencing methods, which are constrained by limited data availability, to innovative single-cell technology that has enabled new AI-based approaches in the omics field. The discussion will provide insights into the structure of the omics data and its inherent complexities, such as sparsity and batch effects. The session will explore key concepts and techniques in developing Single Cell Foundation models, like scGPT. The presentation will also highlight the potential applications of these foundation models in the healthcare industry, particularly in understanding disease mechanisms, tissue composition, and tumor recurrence. The talk is intended for anyone interested in the current state of AI and its potential in the healthcare industry.

Balaji Selvaraj is a Senior Deep Learning Engineer at AstraZeneca, bringing over a decade of expertise in the field. He is recognized as a top-performing participant on Kaggle, ranking in the top 1% and has achieved an exceptional peak global rank of 652 among more than 200,000 data scientists worldwide.

Balaji is proficient in designing and developing advanced deep learning models and pipelines across various data modalities, including images, signals, tabular data, and text. Outside of his professional endeavors, Balaji is passionate about Formula 1, cricket, and engaging in strategic board games.

This talk will discuss the best practices and pitfalls of Conversational User Interface (CUI) design for AI chatbots used in professional settings. Since 2023, Recorded Future has launched several natural language AI features to help our users interact with our complex threat intelligence database—the Intelligence Graph—and aid users in constructing advanced database queries.

For our users, the Recorded Future AI needs to fulfill several use cases: one is to partner with users on an individual level to aid in their iterative knowledge discovery workflows, and another is to produce written short and long-form summaries of user-defined (often esoteric) topics that users then share with stakeholders of varying levels of tech expertise.

Our user base imposes some distinct design challenges. Since our audience consists of cybersecurity professionals, algorithmic transparency is crucial as they place a premium on accuracy, explainability, and trustworthiness. It’s also a heterogeneous audience; every organization’s workflow or environment is unique, and the AI needs to be able to fulfill a wide variety of use cases in a myriad of different socio-technical contexts.

During our presentation, we will share the ways in which we’ve tackled these design challenges and the research that underpins our current design framework.

Björn Berg Marklund has a PhD in informatics and currently conducts UX research focused on understanding cybersecurity professionals’ varied and complex work tasks. Björn’s research examines the challenges of AI adoption from a socio-technical perspective, thus focusing on the “soft” social and cultural factors affecting users’ perceptions and expectations of AI and how these manifest in users’ behaviors in natural language interfaces.

Earth Observation (EO) data is an essential resource for tackling societal, environmental, and industrial challenges. Through initiatives such as the Swedish Space Data Lab (Rymddatalabbet), RISE drives progress in leveraging EO data for both immediate societal benefit and long-term sustainable development.

This talk will delve into RISE's cutting-edge work in integrating artificial intelligence (AI), edge computing, and edge learning to optimize EO data use. Practical examples will highlight projects focused on nature-based solutions for climate change adaptation and mitigation, including biodiversity monitoring, smart agriculture, and urban greening initiatives.

Discover how RISE fosters innovation through collaborative platforms and open calls, driving the development of commercializable solutions from Earth Observation data.

These efforts focus on advancing scalable EO data processing while driving industrial innovation and societal impact. This talk will showcase how RISE and its partners leverage cutting-edge technologies and collaborations to unlock EO data's transformative potential, foster commercialization, and position Sweden as a leader in sustainable innovation and the green transition.

Chiara holds a PhD in Physics and Astronomy from the University of Ferrara. She worked as a researcher in academia for 10+ years. She then worked for two years at the Swedish National Institute for Applied AI (AI Sweden) as a data scientist and lead of space-related projects. Since January 2025, she works at the Research Institutes of Sweden (RISE) as a project manager and researcher at the interface between space and AI.

Defense and the defense industry are in a significant transition. Classical systems are combined with the use of small and cheap systems in large numbers; systems are increasingly connected, and data from a multitude of sources needs to be fused to gain situational awareness; increasingly intelligent decision support systems will allow better and faster decisions. As the reader might guess, AI will play an essential role in the transition. Due to the nature of the domain, requirements for reliable, predictable and ethical systems are essential.

In this talk, we briefly present the domain we are working in and then present a particular example of using LLM to more efficiently find information in huge data. Data are coming from Automatic Identification Systems (AIS) transponders of ships around the coasts, rivers and lakes. In the first step, LLM agents are used to simplify and speed up data filtering by using common language in text or speech. In the second step, it is used for more specific tasks such as identifying smugglers.

Christian is a Software Developer with a passion for pushing technological boundaries. With a background in Computer science, systems and networks at Chalmers University of Technology, Christian is dedicated to advancing Saab's technological capabilities by exploring cutting-edge technologies, such as recent advancements in AI, and promoting their integration into existing systems and products. Currently working in the X Innovation Lab at Saab Surveillance in Gothenburg, he leverages his background in sensor data processing and interest in new technologies to identify new opportunities and drive meaningful advancements.

Recent breakthroughs in AI have unlocked new value for companies to create and capture. Venture capital firms are spraying money around, hoping to back the next AI decacorn. However, it’s far from clear which companies will succeed in both creating and capturing value — i.e. which ones ultimately prove to be financially sustainable.

This talk is about how to operationalize AI as part of a business, not how to develop it. We will explore aspects of building a successful AI business, such as "How do you establish a data moat?", "What are the trade-offs between vertical, domain-specific solutions and more general applications?", "How important is explainability, particularly in regulated industries?", and "How sensitive are users to issues with precision/recall?". The talk will not pretend like there are easy answers but rather discuss what we've learned so far as an industry.

Daniel Langkilde is an experienced machine learning developer and entrepreneur. He started his career as Team Lead for Collection & Analysis at Recorded Future. Since 2018 he is co-founder and CEO of Kognic. Kognic provides the most productive annotation platform for sensor-fusion data. Daniel earned his M.Sc. in Engineering Mathematics at Chalmers University of Technology, where he also served as President of the Student Union and a Member of the Board of Directors. He has been a Visiting Scholar at both MIT and UC Berkeley.

The defence industry faces unprecedented challenges in adapting to evolving threats, and AI-driven solutions are pivotal in addressing these complexities with speed, precision, and scalability. Computer vision is revolutionizing industries by enabling machines to understand and interact with the world through visual data. However, deploying these systems in defense-related environments remains a challenge. 

This talk explores how federated machine learning unlocks new possibilities, enabling decentralized, real-time adaptation of models in dynamic environments. Scaleout is part of a research and development initiative, “AI-based reconnaissance”, addressing the challenge of rapidly changing and unpredictable environments. The project leverages federated learning to enable real-time data collection, analysis, and model fine-tuning while ensuring that raw data remains localized and secure. The demonstration showcases the system’s ability to capture, annotate, and train models within a federated network, efficiently keeping models up-to-date in dynamic environments while operating under constrained network conditions.

David holds a Master’s degree in Engineering Physics, specializing in Scientific Computing and AI. He has done work in various fields, including cybersecurity, computer vision, and fintech. Now a Machine Learning Engineer at Scaleout Systems, David focuses on federated learning projects in the defence sector, tackling challenges in dynamic and secure environments.

Money laundering is a critical enabler of organized crime, integrating illicit profits into the legitimate economy. Despite significant regulatory efforts, the problem is persistent and growing, with recent estimates suggesting a staggering $3.1 trillion of illicit funds flowing globally in 2023. Traditional anti-money laundering (AML) measures are largely based on siloed, rule-based systems, and they struggle to keep up with criminals' sophisticated and ever-changing tactics.

While machine learning offers a promising approach to improve money laundering detection by leveraging more complex patterns in transactional data, significant challenges remain to implement this in practice. These include the lack of realistic, labeled datasets and the complexities of sharing information across institutions.

In this project, we present a holistic AML framework that aims to tackle these challenges; addressing several important issues related to data, collaboration, and explainability. We demonstrate the use of our pipeline to create a synthetic dataset informed by empirical data to simulate realistic inter-bank heterogeneity and show the improved performance from employing federated learning. We also highlight how data quality has a crucial impact on model development and address some important practical concerns around implementing federated learning for AML between different banks.

Edvin is a research engineer at AI Sweden, the national center for applied artificial intelligence in Sweden. Through his work, he contributes to advancing Sweden’s innovation landscape by implementing cutting-edge research in applied projects. His current focus areas include anti-money laundering and energy systems. Edvin is a strong advocate for collaboration through federated and decentralized learning, emphasizing its potential to enhance privacy, robustness, and resilience across industries.

This presentation dives into the role of artificial intelligence (AI) in marketing content production and scaling creativity to new levels both in terms of quantity and quality.

First, we will look at why AI has the potential to completely revolutionize the way we work in marketing, as well as why it is so hard to go from finding the use case to successfully realising its expected value. In this section, we will share a shift in mindset regarding how to perceive AI and the potential it unlocks.

In the main part of the presentation, you will learn about how we:

• Found the right use cases and our framework for prioritisation
• Our framework to avoid developing solutions that are obsolete before they even reach production
• How we organized ourselves to secure long-term success with AI
• The technical architecture of our biggest AI solution
• Examples of content & creatives generated with AI and how

The purpose is to share the mistakes and victories of our AI journey, providing valuable insights & learnings for both engineers and AI decision-makers. While the specific AI use cases are focused on marketing, we believe the learnings and our approach apply to anyone.

Emilie is an Executive Assistant for Group Management at s360. She holds a Master's in Data and Society from the London School of Economics and Political Science. She is part of the team that initially mapped out the relevant use cases, upskilled employees, and secured knowledge sharing in connection with the deployment of AI at s360.

What if collaboration was the key to solving the public sector's most persistent AI challenges?

Jonatan Permert, project manager of a pioneering initiative to build a digital assistant for Sweden's public sector, shares lessons learned from this groundbreaking collaborative effort and provides actionable insights for future AI initiatives.

Spearheaded by AI Sweden, the initiative unites nearly 60 government agencies, cities, and municipalities. Together, they are exploring how a shared Saas solution, built on open-source AI models and a shared technical infrastructure, can help tackle key challenges—like resource constraints, legal compliance, and the lack of sector-specific training data—in AI adoption.

A major milestone of the project is Svea, a prototype digital assistant being tested by over 10,000 Swedish public sector employees. Early results show that over 80% of active users report improved efficiency, quality, and workplace satisfaction.

Perhaps most groundbreaking is the collaboration of over 250 employees from participating organizations working together to annotate data for training common AI models, addressing the challenge of gathering high-quality data tailored to the public sector’s needs.

This presentation delves into the most important challenges and keys to success, highlighting the transformative power of collaboration in public sector AI adoption.

Jonatan Permert is an AI Transformation Strategist at AI Sweden, leading initiatives to accelerate AI adoption across the public sector. Before AI Sweden, he spearheaded AI and innovation projects at Kriminalvården, Sweden's Prison and Probation Service.

With an MSc in Management from Stockholm University and a multidisciplinary background spanning computer science, entrepreneurship, philosophy, and political science, Jonatan brings a holistic perspective to innovation and AI implementation.

Generative AI—particularly Large Language Models (LLMs) and Transformers—rapidly reshapes applications by enabling real-time situational awareness, autonomous decision-making, and efficient data processing. However, conventional cloud-based deployment often proves challenging due to connectivity constraints, cybersecurity risks, and high operational costs.

This talk outlines a systematic approach to deploying advanced generative AI models directly on resource-constrained devices—such as autonomous vehicles, drones, and IoT devices—to ensure autonomy, security, and real-time performance.

We review state-of-the-art system-on-chips (SoCs) from leading manufacturers, examining their capabilities and limitations for executing generative AI workloads. To overcome hardware constraints, we explore model compression techniques such as quantization, pruning, and knowledge distillation, discussing trade-offs between maintaining high-fidelity performance and achieving practical, real-world viability.

Drawing on a recent real-world case study, we present results demonstrating how knowledge distillation and targeted fine-tuning enable SLMs (e.g., Meta’s Llama 3.2) to run on Qualcomm Snapdragon SoCs without requiring massive computational resources.

We will share the engineering techniques and optimizations necessary to adapt next-generation models for resource-constrained environments.

Attendees will learn how to:

• Evaluate hardware accelerators for low-latency, high-throughput inference
• Apply compression methods without sacrificing critical model capabilities
• Balance model size with real-world efficacy
• Leverage emerging trends in SLMs to future-proof AI applications

Jonna Matthiesen is a deep learning researcher specializing in AI optimization for defense, automotive, and IoT applications. With expertise in hardware-aware neural architecture search, model compression, and inference optimization, she focuses on making large language models (LLMs) deployable in resource-constrained environments such as embedded systems and edge devices. She holds a bachelor’s degree in Mathematics from Kiel University, Germany, and a master’s degree in Applied Data Science from Gothenburg University, Sweden. In 2023, Jonna joined Embedl, a company dedicated to efficient deep learning in automotive, defense, and IoT.

Money laundering is a critical enabler of organized crime, integrating illicit profits into the legitimate economy. Despite significant regulatory efforts, the problem is persistent and growing, with recent estimates suggesting a staggering $3.1 trillion of illicit funds flowing globally in 2023. Traditional anti-money laundering (AML) measures are largely based on siloed, rule-based systems, and they struggle to keep up with criminals' sophisticated and ever-changing tactics.

While machine learning offers a promising approach to improve money laundering detection by leveraging more complex patterns in transactional data, significant challenges remain to implement this in practice. These include the lack of realistic, labeled datasets and the complexities of sharing information across institutions.

In this project, we present a holistic AML framework that aims to tackle these challenges; addressing several important issues related to data, collaboration, and explainability. We demonstrate the use of our pipeline to create a synthetic dataset informed by empirical data to simulate realistic inter-bank heterogeneity and show the improved performance from employing federated learning. We also highlight how data quality has a crucial impact on model development and address some important practical concerns around implementing federated learning for AML between different banks.

Kristiina is a data scientist at Handelsbanken, working in the Advanced Analytics & AI team. In this role, she applies machine learning and AI to build innovative products and services and make the bank more data-driven. She is also involved in research collaborations exploring new technologies, including a joint project with AI Sweden and Swedbank to use federated learning to improve money laundering detection. Kristiina holds a PhD in scientific computing from Uppsala University, where her research focused on machine learning, statistical computing, and distributed systems.

Intelligence is inherently corporeal—the value functions of why we turn towards what is good (food! sex!) and away from what is bad (harm! death!) are deeply rooted in our embodied existence: our physical, emotional, and somatic ways of being in the world. Consequently, aesthetics and ethics, too, are corporeal, grounded in the lived realities of our bodily existence. The current wave of AI exists predominantly in the realm of language. While this realm is not entirely detached from materialities such as electricity and servers, it operates largely without the visceral stakes of life, living, and death that define our embodied experience. But what happens when AI begins to bridge this gap, embedding itself in and interacting with our corporeal bodies? How might this shift redefine the aesthetic and ethical implications of AI in our lives? Through exploring designs that come into close contact with our bodies, this talk invites a discussion on intelligence, meaning-making, aesthetics, ethics, and how we can design AI to support a good and meaningful life.

Kristina “Kia” Höök is a Professor of Interaction Design at the KTH Royal Institute of Technology, Sweden, and works part-time at RISE. She is known for her work on designing for bodily engagement in interaction through soma design. She has obtained numerous national and international grants, awards, and fellowships, including the Cor Baayen Fellowship by ERCIM and the INGVAR award. She is an ACM Distinguished Scientist, elected to the ACM SIGCHI Academy, and was recently awarded the ACM SIGCHI Lifetime Service Award. Höök is a horseback rider, mother, grandmother, and feminist.

This talk explores the disconnect between MLOps fundamental principles and their practical application in designing, operating and maintaining machine learning pipelines. We’ll break down these principles, examine their influence on pipeline architecture, and conclude with a straightforward, vendor-agnostic mind-map, offering a roadmap to build resilient MLOps systems for any project or technology stack.

Former Creative Director with extended expertise in product and strategy, Lex Avstreikh now works as the Head of Strategy at Hopsworks; a Swedish startup at the forefront of machine learning infrastructure. He focuses on identifying pivotal market trends and executing strategic initiatives that secure and advance Hopsworks’ position as a global leader in the ML industry.

A goal in the Swedish healthcare system is that every cancer patient should receive a personalized care plan (Min Vårdplan) - a comprehensive document that serves as a national knowledge support system. Currently, these plans are manually created by a team of healthcare professionals, including contact nurses, doctors and physiotherapists, consuming valuable time in an already resource-constrained healthcare environment.

Tenfifty, together with Regional Cancer Centers of Sweden, are developing a solution to allow healthcare professionals to spend less time on the tedious tasks involved in creating the care plan and instead let them direct their expertise to the complex aspects. The goal is that they should be reviewers instead of writers. Our solution uses an agentic workflow where LLMs are tasked with decisions about document structure, adaptation of existing content, generating new material and providing references where these are needed.

This talk will detail our technical implementation and the challenges we faced. It is aimed at anyone working in the generative AI space, and especially those who are interested in using LLMs in multi-step workflows.

Linus earned his master's degree in computer science from Chalmers University of Technology in 2023 and has since been working as a data scientist at Tenfifty. He works with implementing AI solutions across various business domains. Having completed his master's thesis in the field of NLP, he has developed a particular interest in NLP-based projects and their broad range of applications.

"AI Governance in a Changing Regulatory Environment" emphasizes the importance of responsible, ethical, and transparent AI development. It provides an overview of global AI regulations, focusing on key regions and markets, and discusses the impact of these regulations on AI development. The presentation highlights the challenges in AI governance, including identifying high-risk AI systems, compliance difficulties, and ethical considerations. It also explores the opportunities that robust AI governance frameworks offer. The presentation concludes with actionable strategies for navigating regulatory complexities and best practices for implementing AI governance. It is an essential resource for professionals seeking to ensure their AI solutions are compliant and cutting-edge.

Luis Martínez is dedicated to ensuring the ethical and compliant development and deployment of AI solutions. He is an expert in ICT and regulatory affairs, currently serving as the Global AI Compliance Manager at Assa Abloy. He holds a PhD in ICT from the Royal Institute of Technology (KTH) in Sweden and a Master's in Telecommunications Engineering and Systems Engineering from the National University of Colombia. Luis has worked at Ericsson and Volvo Cars, excelling in AI governance and industrial standards (5G).

As the global energy transition accelerates, grid flexibility has emerged as a critical enabler for decarbonizing electricity systems worldwide. However, unlocking this grid flexibility requires accurate, high-resolution forecasting of future global grid conditions. To meet this challenge, Electricity Maps has developed a robust, interconnected machine-learning platform that can predict the future state of electricity networks worldwide. Our approach leverages thousands of specialized ML models—each trained on granular data from a specific grid—which are then interconnected to account for the real-world linkages between cross-border electricity networks.

Marcus Garsdal is a MLOps Engineer within the Grid Forecasts team of Electricity Maps, orchestrating the lifecycle of the thousands of machine Learning models used by a forecasting engine that predicts the future state of electricity grids worldwide. Relying on his previous experience working at a Danish energy and weather forecasting company, delivering operational forecasts to +100GW of renewable capacity, as well as his academic knowledge of sustainable energy topics and machine learning, Marcus is a driving force for developing the grid and carbon data necessary to unlock the large-scale flexibility required for ensuring a greener future.

This talk will give introductory information about time series data forecasting and examples of its applications, and then delve into some lessons learnt from working on many time series forecasting topics, examples of the lessons:

1. Garbage In, Garbage Out: Emphasizing the critical need for investing in data quality improvement.
2. Investing in Exploratory Data Analysis always pays off: The importance of thoroughly understanding data before modelling, using techniques like autocorrelation and decomposition.
3. Model Size Doesn't Necessarily Matter: Demonstrating how advanced AI methods can sometimes underperform compared to simpler models.
4. Think Bigger: Exploring the benefits of examining relationships between multiple time series for more accurate and cost-effective models.
5. Continuous Monitoring Challenges: Addressing the difficulties of evaluating forecasting models with traditional monitoring techniques in highly seasonal data.

These insights aim to provide a comprehensive understanding of the practical aspects and challenges of time series forecasting, offering valuable takeaways for industry professionals.

Mohamed is a data scientist with 10+ years of experience in applying data analytics to solving business challenges, he is part of the Discipline Expert community at Siemens Energy and is leading the Time Series Analytics R&D program.

This session explores how to use Python and GPT to build real-world NLP applications. Structured in four key parts, it begins with foundational concepts, advances into NLP fundamentals, explores applications of GPT, and concludes with a perspective on AI’s roadmap. Attendees will gain insights into the GPT family tree, enterprise architecture and NLP applications in manufacturing, sales, marketing and legal. This talk is designed for IT professionals, data scientists, ML engineers, and entrepreneurs.

Neil is a Data Scientist with experience from CIO, CTO and Enterprise Architect roles. He strongly believes that businesses have two (and only two) basic functions: Marketing & Innovation. He helps firms create and retain customers through his expertise in numerical analysis, computer vision, natural language processing, IoT and digital twins. His home turf is Helsingborg, Edinburgh, and London. He serves clients around the world. Neil was awarded a PhD from Heriot-Watt University’s Subsea Robotics Lab and is the author of the Python GPT Cookbook.

Earth Observation (EO) data is an essential resource for tackling societal, environmental, and industrial challenges. Through initiatives such as the Swedish Space Data Lab (Rymddatalabbet), RISE drives progress in leveraging EO data for both immediate societal benefit and long-term sustainable development.

This talk will delve into RISE's cutting-edge work in integrating artificial intelligence (AI), edge computing, and edge learning to optimize EO data use. Practical examples will highlight projects focused on nature-based solutions for climate change adaptation and mitigation, including biodiversity monitoring, smart agriculture, and urban greening initiatives.

Discover how RISE fosters innovation through collaborative platforms and open calls, driving the development of commercializable solutions from Earth Observation data.

These efforts focus on advancing scalable EO data processing while driving industrial innovation and societal impact. This talk will showcase how RISE and its partners leverage cutting-edge technologies and collaborations to unlock EO data's transformative potential, foster commercialization, and position Sweden as a leader in sustainable innovation and the green transition.

Olof Mogren is the director of deep learning research at RISE, co-founder of Climate AI Nordics, and co-principal investigator for CLIMES, the Swedish Center for Impacts of Climate Extremes. Holding a PhD in machine learning from Chalmers University of Technology (2018), he focuses on both foundational and applied machine learning. His work spans areas such as computer vision and soundscape analysis, with a strong emphasis on climate change adaptation and environmental monitoring. Recent project topics include biodiversity monitoring, efficient distributed machine learning, remote sensing, stream flow forecasting, and smart fire detection using machine listening.

Data work has evolved significantly in the past decade, leaving traditional orchestrators struggling to keep pace. In this talk, I’ll explore five key shifts reshaping how data teams approach orchestration:

1. Standardization: Modern tools like dbt and Fivetran require orchestrators to go beyond running custom scripts, focusing instead on seamless tool integration and clear data lineage.
2. Less DAG Code: Reference-based scheduling has reduced the need for complex orchestration code, simplifying big DAG management.
3. Reactive Scheduling: Real-world data changes dynamically, requiring jobs to run when data arrives, enabling low-latency pipelines without streaming complexity.
4. Data Products: As data powers customer-facing and business-critical products, traditional analytics-focused tools fall short of essential software engineering practices like testing and CI/CD.
5. Bundling: Companies are consolidating their tool stacks to cut costs, reduce complexity, and streamline workflows.

I will discuss how to design orchestration systems that embrace these trends, regardless of the tools you use, and share how we’re building Twirl to support these shifts natively.

Rebecka Storm has a background in machine learning and has held data leadership roles as ML Lead at iZettle and Head of Data at Tink. She is now co-founder and CPO of data orchestration startup Twirl, where she is dedicated to simplifying how data applications are built.

In 2018, Rebecka co-founded Women in Data Science Sweden, an organization that promotes inclusivity through conferences, mentorship programs, a speaker database and more to inspire and support women working in data.

This talk will discuss the best practices and pitfalls of Conversational User Interface (CUI) design for AI chatbots used in professional settings. Since 2023, Recorded Future has launched several natural language AI features to help our users interact with our complex threat intelligence database—the Intelligence Graph—and aid users in constructing advanced database queries.

For our users, the Recorded Future AI needs to fulfill several use cases: one is to partner with users on an individual level to aid in their iterative knowledge discovery workflows, and another is to produce written short and long-form summaries of user-defined (often esoteric) topics that users then share with stakeholders of varying levels of tech expertise.

Our user base imposes some distinct design challenges. Since our audience consists of cybersecurity professionals, algorithmic transparency is crucial as they place a premium on accuracy, explainability, and trustworthiness. It’s also a heterogeneous audience; every organization’s workflow or environment is unique, and the AI needs to be able to fulfill a wide variety of use cases in a myriad of different socio-technical contexts.

During our presentation, we will share the ways in which we’ve tackled these design challenges and the research that underpins our current design framework.

Ruxandra Teodoru is a product designer with extensive experience building user-focused products and managing high-performing teams. Currently a Product Design Director at Recorded Future, she specializes in strategic design, AI, and fostering innovation to deliver impactful solutions that bridge business goals and user needs.

Simple analysis tools remain pivotal for success in the fast-developing world of AI. More than not, even basic methods and lightweight tools in a Data Scientist’s toolbox can lay the foundation for magical outcomes.

We present a success story from the real estate sector, where we developed an AI-powered solution to optimize energy systems in apartment buildings. In its first year of application in more than 3000 apartments across Sweden, we achieved a significant 10–15% reduction in total energy consumption, saving more than 180 tons of CO2 equivalent. Simultaneously, temperature stability for tenants was improved, and the day-to-day workload for technicians and administrators was reduced.

While the developed product now incorporates machine learning models and AI-driven time-series analysis, the journey began with the most fundamental tools in the shed.
Leveraging cutting-edge AI innovations on the one hand and deriving insights from linear regressions on the other are not mutually exclusive—they are complementary steps toward creating impactful solutions.

So, think big, even if you start small!

Dr Steffen Therre has worked as a Data Scientist since 2022, with a focus on climate-smart solutions, energy efficiency in the real estate sector, and developing optimization models. As a consultant in the AI and Data Science environment, he supports clients in building and adopting intelligent products to improve their business. Steffen is dedicated to advancing a climate-neutral future through the effective use of modern tools. He holds a PhD in physics from Heidelberg University, where his climate research even took him into the caves beneath the Caribbean rainforest.

Agentic systems have demonstrated their ability to leverage foundation models (FMs) to tackle complex, real-world challenges. With multi-agent collaboration to orchestrate multiple agents, these systems enable autonomous collaboration, decision-making, and efficient problem-solving across diverse environments and industries. 

In this talk, we will take a first-principles approach to deconstruct the architecture of multi-agent collaboration frameworks via industry use cases and explore how specialized agents could be prompted and orchestrated to boost intra-agent collaboration. We’ll touch on methods for evaluating these multi-agent systems and how to deploy and scale these systems in production in a governed and responsible manner.

Tingyi Li is an Enterprise solutions architect at Amazon Web Services, where she leads strategic engagements with major Nordic enterprises on their cloud-optimized digital transformations. As founder and leader of AWS Nordics Generative AI community, she is dedicated to Generative AI productionisation, working towards enterprise adoption and governance of industry-specific GenAI solutions and agentic workflows. Prior to AWS, she worked as a Data & AI Engineer at Intel, Foxconn, and Huawei, building large-scale intelligent industrial information platforms and data integration systems with advanced data pipelining and AI/ML technologies.

Tingyi is a frequent speaker at premier conferences globally, including AWS Re:Invent, AWS Summit, QCon, TDC, NextM, Nordics AI Week, IEEE Cloud Summit, IEEE WIE Leadership Summit, etc. and is the featured instructor for the flagship GenAI courses at the University of Oxford. In her spare time, she works as a part-time illustrator who writes novels and plays the piano.

Zenseact is advancing towards an autonomous driving (AD) stack based on deep learning (DL) models that perform end-to-end sensor and temporal fusion. Training these models demands a significant computational budget that scales with in-vehicle compute capacity. DL models tend to only operate safely within the domain of their training data. Meanwhile, annotating multi-second data sequences is vastly more expensive than single images, posing challenges to expanding the operational domain of next-generation AD systems.

Manual annotation is insufficient to meet these demands. Pseudo-annotations —algorithmically generated annotations— offer a solution, as they are not constrained by real-time or embedded hardware requirements. This enables the use of very large models. To train such models without increasing annotation needs, we employ self-supervised learning, training models to understand relationships in input data. Scaling this approach to large datasets and model sizes produces systems highly capable of solving downstream tasks. In summary, by increasingly leveraging a large compute capacity, the human annotation effort can be offloaded.

Ensuring the safety of DL models is another challenge that might only be tackled by simulating sensor data to evaluate performance under adverse conditions. Neural Radiance Fields (NeRFs) and 3D Gaussian splatting are emerging as critical technologies for enabling and scaling these simulations.

I completed my PhD in particle physics at CERN's CMS experiment, focusing on leveraging deep learning to enhance searches for new physics. Seeking a career with real-world impact, I transitioned to deep learning engineering at Zenseact, initially specializing in object detection. Later on, I helped build the foundation for Zenseact's next-generation deep learning models that do sensor and temporal fusion. This transition led to the creation of a new "Deep Learning Data Enrichment" area, which I now lead. This area aims to automate annotation processes with large offline models and mine high-value data.

Retrieval-augmented generation (RAG) systems integrate external information sources with advanced large language models (LLMs), promising more factual and context-aware outputs. However, assessing their performance remains a significant challenge—it involves assessing not only generated text quality but also the relevance and correctness of retrieved information. Recent developments have introduced new frameworks and metrics, but the field has not converged on widely accepted standards for robust evaluation.

In this talk, we explore current approaches to RAG evaluation and discuss key open questions from ongoing research in this area. Our goal is to encourage more systematic approaches for measuring and comparing RAG performance in a continually advancing field.

Yue is a Machine Learning (ML) Engineer at Modulai, an ML consultancy company in Sweden. There she has worked on various projects in the healthcare, legal and finance sectors. Before joining Modulai, Yue’s PhD research at KTH focused on applying AI models to breast cancer risk assessment and detection in mammograms. Earlier, she pursued her Master’s in Computer Science at KTH, Sweden and TU Delft, the Netherlands.

As the global energy transition accelerates, grid flexibility has emerged as a critical enabler for decarbonizing electricity systems worldwide. However, unlocking this grid flexibility requires accurate, high-resolution forecasting of future global grid conditions. To meet this challenge, Electricity Maps has developed a robust, interconnected machine-learning platform that can predict the future state of electricity networks worldwide. Our approach leverages thousands of specialized ML models—each trained on granular data from a specific grid—which are then interconnected to account for the real-world linkages between cross-border electricity networks.

Íngrid Munné Collado is the Technical Lead and a Senior Machine Learning Engineer in the Grid Forecasts team at Electricity Maps, driving the development of scalable forecasting systems for power and carbon data. With a Ph.D. in Electrical Engineering and experience in electricity markets, demand response, and renewable energy forecasting, she contributes to developing forecasting models that enhance flexibility and accelerate the decarbonization of the electrical grid.