The 2026 GAIA Conference

At Jeppesen, we develop world-leading optimization solutions for the aviation industry, solving NP-hard problems like crew scheduling using methods such as Linear Programming and Mixed-Integer Programming.

In this talk, we will share insights from a recent exploration where our AI enablement and optimization teams collaborated to evaluate the potential of combining Evolutionary Algorithms with the creative power of Generative AI in our problem space. During this exploration, we pushed the boundaries of traditional optimization by enabling AI to mutate algorithms and code using biologically inspired algorithms. This fusion of "old meets new" not only sparked innovative solutions but also demonstrated the value of thinking outside the box when tackling optimization challenges and gave a taste of what the future could hold.

Alexander is a Data Scientist at Jeppesen ForeFlight, where he focuses on crew scheduling optimization for the aviation industry. His technical background is rooted in C++, and he currently explores the synergy between traditional optimization and AI. Specifically, he has been looking into how evolutionary algorithms, such as OpenEvolve, can be integrated into the optimization process. At GAIA, he will discuss how we can bridge these methods to solve complex scheduling challenges.

In this talk, we explore the potential of generative AI, specifically Large Language Models (LLMs) and Vision-Language Models (VLMs), to support the development and operation of safe autonomous driving.

The process begins by identifying critical scenarios in real-world data, then generating safety concepts to mitigate them, and ultimately implementing them in code. To enable this "from concept to code" workflow, we first examine the strengths and limitations of LLMs in automotive system engineering tasks. Building on these insights, we design multi-agent systems that enhance robustness, consistency, and validity of the generated artifacts. We then showcase a set of tailored pipelines, each supporting a different activity in the DevSafeOps cycle, from field data to code generation, refinement, and evaluation. This approach is not limited to rule-based software; it also extends to ML-based software units, enabled by integrating 3D Gaussian Splatting for synthetic data generation and VLMs to trigger data collection.

These AI-enabled pipelines accelerate automotive system and software engineering activities, illustrate how generative AI can integrate into the continuous software development process in automotive, and highlight the emerging role of LLMs in this field.

Ali Nouri is an AI expert at Volvo Cars, working on autonomous driving, and a researcher at Chalmers University of Technology. He has more than ten years of experience in autonomous driving systems. He represents Sweden in international ISO standardization efforts, including ISO 8800 (Safety of AI). His research focuses on accelerating DevOps for autonomous driving software development through generative AI.

This talk explores what happens when AI systems are designed not only to respond to our emotions but to influence them. Emotional AI tools, including chatbots and AI companions, increasingly use affective design, reward loops and adaptive emotional cues that can quietly shape user behavior. These techniques, grounded in behavioral psychology and neuroeconomics, raise new questions about autonomy, manipulation and the limits of traditional consent.

Ana Catarina de Alencar shows how emotionally responsive AI can trigger patterns of dependency, create subtle forms of influence and blur the boundary between persuasion and manipulation. When an AI system comforts, validates, or rewards users at precisely the right emotional moment, can we still say that consent is freely given? And what does informed consent mean when the system itself affects the user’s emotional state and decision-making?

Drawing from emerging real-world cases (including situations of emotional attachment to AI companions), the talk highlights why existing privacy and consent frameworks are insufficient for emotionally interactive technologies. Attendees will gain a clear understanding of how manipulative design manifests in Emotional AI, why it matters for ethics and regulation, and what needs to change to protect user autonomy.

Ana Catarina de Alencar is an international lawyer and ethicist based in Paris, working at the intersection of AI governance, regulation, and philosophy. She is the Resident Philosopher at The AI Collective, a San Francisco–based organization dedicated to shaping responsible AI futures. Ana is also a PhD researcher at Université de Lille, where she develops an interdisciplinary project on emotional AI, bridging law, philosophy, and neurobiology. She holds a Master’s degree in Philosophy and Technology of Law and is the author of several publications, including Artificial Intelligence, Ethics and Law (2022).

This keynote explores the intersection of visualization, artificial intelligence, and human decision support, focusing on how multimodal interfaces are shaping next-generation AI-enabled workflows. Effective visualization empowers users to interpret complex data, while AI augments human insight, making decision processes more robust and transparent. Multimodal interfaces—combining visual, auditory, and tactile channels—are emerging as pivotal tools in facilitating seamless human-AI collaboration, enabling users to interact with systems in more intuitive and meaningful ways.

To realize these advancements, Sweden requires a robust AI compute infrastructure capable of supporting both research and industrial innovation. The talk will highlight examples such as the build-up of Sferical AI to provide key industries in Sweden with state-of-the-art AI compute. The Wallenberg AI, Autonomous Systems and Software Program (WASP) plays a central role in this landscape, driving forward the development of advanced AI technologies and cultivating a new generation of experts. Through its comprehensive initiatives, WASP provides critical competence that bridges the gap between academic breakthroughs and real-world industrial adoption.

Anders Ynnerman is a renowned Swedish scientist specializing in scientific visualization, computer graphics, and visual AI. He currently serves as the Director of Strategic Research at the Knut and Alice Wallenberg Foundation and is Executive Chairman of Spherical AI. With a strong background in supercomputing infrastructure, Ynnerman has served as director of the National Supercomputer Centre (NSC), contributing significantly to the development of Sweden’s computational research capabilities. In addition, he has played a pivotal role in the Wallenberg AI, Autonomous Systems and Software Program (WASP)—having previously served as Program Director and now acting as Chair of WASP, Sweden’s largest research program. Ynnerman is also the leader of the national WISDOME project for Science Communication. His research base is as a Professor of Scientific Visualization at Linköping University, where he leads the Visualization Center C. Through these leadership positions, Ynnerman continues to advance the field of digital science communication, supercomputing, and innovative research in Sweden and beyond.

This presentation introduces the new Guidance Note NI692 developed by Bureau Veritas Marine & Offshore for assessing Machine Learning Systems (aaccessible here).

This Guidance Note provides recommendations for the transparent and trustworthy development and operation of machine learning systems in marine settings. It emphasizes human oversight and risk-based assessment across the system lifecycle, from data and design to deployment, monitoring, and maintenance.

Key highlights of the Guidance Note include:

• Operational context: description of the ML system's objectives, functions, operational and environmental boundaries (Concept of Operations, Operational Envelope, and Operational Design Domain).
• Functional analysis: identification and description of specific functions and sub-functions relying on the ML system.
• Human oversight: defining system automation and corresponding human supervision, roles and responsibilities, and human-machine interactions.
• Risk management: strategies for risk assessment and mitigation.
• Data governance: ensuring data quality, integrity, and security through data collection and preprocessing.
• Model development: recommendations for testing and validating ML models.
• Operational governance: continuous monitoring and maintenance protocols during operation.

NI692 builds upon the EU AI Act, ISO standards, and the IMO MASS Code, and emphasizes key aspects of ML systems such as transparency, explainability, human oversight, and traceability.

Bérénice Le Glouanec holds a master's degree in Language Technology from the University of Gothenburg. She is the AI and data technical expert at Bureau Veritas Marine & Offshore, working within the Digital and Autonomous Ship team in the Rule Development department. She is responsible for drafting rule notes and recommendations, including NI692, a guideline that addresses the entire lifecycle of machine learning systems and ensures their trustworthiness. She also contributes to the International Association of Classification Societies (IACS) on the revision of Recommendation 183 on Ship Data Quality and follows AI standardization activities within AFNOR and CEN-CENELEC.

In this talk, The Good AI Lab and Doctors Without Borders share how they are building a long-term collaboration where AI is taught and developed openly, accountably, and in service of humanitarian work. They will outline how AI enablement, strategy, and applied research projects can fit into a single roadmap, and what they have learned so far about making AI useful in high-stakes, low-resource settings without losing sight of dignity, context, and constraints.

Carlo Rapisarda is a software engineer and co-founder of The Good AI Lab. He currently works at GeoGuessr, focusing on product engineering with an emphasis on mobile development and 3D/graphics experiences. Previously, he led applied AI initiatives at Framna. At The Good AI Lab, Carlo helps translate responsible-AI principles into deployable systems and practical training programs for humanitarian and public-interest partners, helping ensure that real-world AI deployments are useful, safe, and aligned with mission needs.

Marketing teams rarely pursue a single objective: lifting sales must coexist with brand‑awareness targets, customer‑acquisition caps, or sustainability constraints. Traditional media‑mix models (MMMs) treat each metric in isolation, leaving decision‑makers to reconcile incompatible recommendations by hand. This talk demonstrates how graphical Bayesian modelling enables simultaneous inference and optimization across multiple objectives, producing allocation strategies that respect every KPI—and clearly flag when trade‑offs become infeasible.

Description: During the days when every euro of media spend juggles revenue, reach, churn, and business constraints, classic media‑mix models fall short: they optimize one KPI and leave the rest to managerial guesswork. This talk reveals how graphical Bayesian modelling lets multiple causal MMMs—each focused on a different target variable—cohabit in a single, principled budget‑allocation problem. 

We'll focus on the PyMC Ecosystem, showing how PyTensor, PyMC and PyMC-Marketing help to solve this issue. Mentioning the advantages of graphical models, and how to use them to build causal media mix models and perform complex operations in a very straightforward manner.

Attendees will leave with a reproducible notebook template, a principled framework for embedding several MMMs in one optimisation problem, and a checklist for detecting—and communicating—when certain goals cannot be met concurrently. The material assumes familiarity with Bayesian inference but will provide a concise refresher on PyMC syntax and the latest PyMC‑Marketing utilities.

Carlos Trujillo is a marketing scientist and data scientist specializing in Bayesian statistics and advanced analytics. He has over seven years of experience applying statistical modeling and machine learning to real-world marketing problems, including work at Omnicom Media Group and Bolt, and he currently works at Wise. He is a contributor to PyMC-Marketing, an open-source Bayesian marketing analytics library within the PyMC ecosystem, where he helps develop tools for media mix modeling, budget optimization, and causal inference. His work bridges theory and practice, and he regularly collaborates with international teams while contributing to the open-source community and technical knowledge sharing.

This case concerns the development of an end-to-end geospatial pipeline that integrates curated spatial attributes with Husqvarna’s internal data to produce a coherent view of golf course environments. Developed in collaboration with Knowit Solutions, the pipeline establishes the technical foundation for geospatial analytics that drive business insights. Structured spatial indexing and robust metadata contribute to an annotated dataset that supports computer vision workflows in validating patterns and enriching geospatial analytics.

Husqvarna is further advancing its data-driven approach by using geospatial analytics to generate deeper insights into green spaces and understand regional variation. This evolving technical foundation is strengthening Husqvarna’s ability to combine visual and spatial information for more precise green space evaluation and identification, enabling more focused and informed decision-making across the organization.

Christos Marinos is a Data Engineer specializing in designing and building scalable, robust data platforms that enable analytics, AI, and machine learning products. He focuses on data extraction, transformation, and modeling, delivering reliable pipelines that support end-to-end AI and ML workflows. He is dedicated to creating data products that generate business value, working closely with stakeholders to align technical solutions with real needs. With a strong emphasis on data governance, scalability, and operational robustness, he builds maintainable platforms that support long-term, value-driven decision-making.

Unlabeled data is abundant in industry, but ground truth is scarce. Trafikverket captures hundreds of terabytes of high-resolution imagery annually to monitor Sweden’s railway network. The sheer volume of this data renders manual labeling—and therefore traditional supervised learning—unfeasible.

This talk explores how we are overcoming the "labeling bottleneck" by deploying a self-supervised foundation model trained on this massive, unlabeled archive. We will bridge the gap between theory and practice, starting with a comparison of Self-Supervised Learning (SSL) paradigms. Moving beyond theory, we will demonstrate the practical application of SSL in a large-scale industrial setting.

Attendees will learn how Trafikverket leverages SSL to learn robust visual representations, enabling the fine-tuning of models for critical downstream tasks—such as detecting rail cracks and identifying key infrastructure assets—with minimal labeled data.

Edvin Listo Zec is a Senior Machine Learning Engineer at Eghed, where he applies advanced deep learning to critical business challenges. He holds a PhD from KTH Royal Institute of Technology, where his research focused on distributed deep learning. Previously, Edvin served as a Research Scientist at RISE Research Institutes of Sweden and a Visiting Researcher at NYU. With a background spanning representation learning and out-of-distribution generalization, he is dedicated to bridging the gap between theoretical research and impactful, sustainable applications in the physical world.

The talk will outline how Zenseact is using spatio-temporal Transformers to realize E2E driving with safety guardrails. We'll discuss opportunities and challenges that come with this approach.

Erik is currently Senior Director of AI and Perception at Zenseact. Prior to this role, he held several key technical leadership positions at Zenseact (and formerly Zenuity), including Chief AI Officer, Chief Architect, Product Area Owner for Computer Vision, and Technical Expert in Deep Learning.

From 2014 to 2017, he served as Director of Automated Driving and Preventive Safety at Autoliv’s global research division. Erik holds a PhD in superstring theory from Chalmers University of Technology (2006) and has eight years of experience in statistical accident research at Autoliv.

In the depths of a Boliden mine, 700 meters underground, running traditional cloud-first AI on heavy-duty machinery falls short on delivering results. In a collaboration with Volvo Group and academic partners, we validated a novel approach: turning the trucks themselves into intelligent and interactively queryable computational databases. By embedding a tiny combined main-memory database and computation engine directly onto heavy-duty mining vehicles, we transformed the fleet into a distributed system where data streams are analyzed and queried at the source in real-time.

This talk shares the architectural lessons learned from deploying this "database-on-wheels" model to monitor critical metrics like battery health, energy regeneration, and driving patterns in a connectivity-constrained environment.

Beyond the immediate deployment, this architecture offers a fundamental shift in how we build industrial AI. We will explore how exposing physical assets via a familiar SQL-like interface paves the way for the next generation of Agentic Workflows. Instead of dealing with rigid firmware cycles, autonomous agents can simply “query” the fleet for insights or update model weights as easily as updating a row in a database table. We will discuss the trade-offs of edge-native computational query processing and how this approach decouples rapid AI innovation from the slower engineering cycles of heavy machinery.

Erik Zeitler is a co-founder of Stream Analyze and a database systems engineer focused on real-time analytics on constrained edge devices. He holds a Ph.D. in Computer Science from Uppsala University, with research contributions to large-scale data stream processing. Before founding Stream Analyze, he led data infrastructure architecture at Klarna, enabling real-time risk and fraud systems at scale. Erik’s work spans academia and industry, with deployments ranging from cloud platforms to heavy-duty industrial vehicles.

At Jeppesen, we develop world-leading optimization solutions for the aviation industry, solving NP-hard problems like crew scheduling using methods such as Linear Programming and Mixed-Integer Programming.

In this talk, we will share insights from a recent exploration where our AI enablement and optimization teams collaborated to evaluate the potential of combining Evolutionary Algorithms with the creative power of Generative AI in our problem space. During this exploration, we pushed the boundaries of traditional optimization by enabling AI to mutate algorithms and code using biologically inspired algorithms. This fusion of "old meets new" not only sparked innovative solutions but also demonstrated the value of thinking outside the box when tackling optimization challenges and gave a taste of what the future could hold.

Jacob is a 27-year-old data scientist at Jeppesen ForeFlight, where he works as part of the AI Enablement Team. Their mission is to elevate the organization's overall AI maturity, with mainly focusing on assisting developer teams in overcoming both immediate and long-term challenges.

As a pilot in his free time, he is passionate about blending his love for aviation with his burning interest in AI. Every day, he strives to explore how we can leverage predictive technologies to make smarter, more informed decisions in the aviation industry.

While LLMs and deep learning are all the rage nowadays, many problems still make some of these models infeasible. In cases such as network filtering, real-time security monitoring, etc., where one might need to make a decision within a few nanoseconds, there is no time to copy data to a userspace process, let alone to a GPU. This talk will outline work on building ML models to run in Linux kernel space, in particular, the eBPF virtual machine and the restrictions imposed on running models there. We will use a fairly naive dataset from Kaggle for malicious traffic detection and showcase how a model can be trained, compiled, and deployed in a real-life kernel without restarting the machine.

eBPF in and of itself poses a set of interesting constraints:

1. No FPU (no floating point arithmetic that drives many models today)
2. No heap(-ish), keeping weights etc. becomes a bit complicated
3. No stdint/stdlib, normal stdlib features (like memset/malloc) you would expect do not exist, making C-compatibility harder.

What you get in return is a formally-verified program (and model) that is guaranteed to be safe w.r.t., e.g. out-of-bound reads, predictable runtime complexity, and more importantly, a program you can deploy on any Linux machine with a fairly small set of capabilities without having to modify the running kernel.

The talk is aimed at both ML Engineers, developers and a little bit of SecOps, where I will outline the task and dataset with some flashbacks of caveats in traditional feature-engineering, in particular for (network) packet inspection. There might be lessons on how NOT to build datasets and on the importance of domain expertise in both dataset creation and feature engineering. Additionally, Jesper will give an introduction to eBPF and kernel development, and share lessons learned from building code (and models) that is portable across eBPF while still maintaining backwards compatibility with the original embedded-C targets.

In some sense, this is an ode to classical ML and to how traditional feature engineering still has a place, even if it's a small one.

This work is not sponsored by any company; it is open-source and in the public domain.

Jesper is a freelancing Tech Lead and Senior Machine Learning Engineer with a career spanning through AI/ML, software development, architecture, and research. His journey in machine learning spans more than 20 years, from developing an embedded computer-vision system to more recently internet-scale infrastructure, where roughly 30% of internet traffic flows through his models. Over the years, his career has spanned everything from natural-language processing to reinforcement learning (and everything in between). He is also an active contributor to many open-source projects, including the Linux kernel.

We are building AI systems everywhere—from customer service to clinical decision support—while not even fully trusting the chatbots. And we are wise not to trust them. The truth is that when we talk about "human-in-the-loop", we are giving our AI an F, and we are losing out on most of the automation and scalability. So if we can't trust a simple chatbot, how can we design enterprise-wide autonomous AI?

This talk dives into the root cause of those trust issues: the gap between probabilistic-based prediction and fact-based understanding, and why modern AI needs more than just bigger models doing the RAG time shake to become reliable.

I will dig into the concept of World Models: internal representations of reality. Every intelligent system—whether a worm, a human, or an AI—must build a small world inside itself. This internal model is what allows it to predict rather than just react. It’s how the brain, the worm, and the algorithm navigate and find their way in a messy world.

Through examples from various industries (including modelling global supply chains with millions of graph relationships), we’ll explore how grounding AI in structured knowledge dramatically improves its ability to reason, stay consistent, and avoid hallucination.

Attendees will learn how knowledge graphs, ontologies, and hybrid neuro-symbolic architectures create AI systems that can explain their answers, keep track of facts, and align with organisational logic. Finally, we’ll look at practical patterns for bringing reliability into real-world AI setups—from digital twins and decision support to enterprise agents that can be trusted with actual work.

If your chatbot also has “trust issues,” this talk will help you understand why and what it takes to fix the relationship.

Johan Müllern-Aspegren is Emerging Tech Lead at Capgemini’s Applied Innovation Exchange Nordics. With a passion for innovation and shaking things up, he co-founded Imagine Scandinavia’s largest bootcamp for innovation, founded the Careful AI Test Lab in the City of Helsingborg, and established RIOT Labs at Capgemini to churn out bleeding-edge prototypes. Now part of the AI Futures Lab, he works hands-on with tomorrow’s AI capabilities while helping organisations build AI that is reliable, understandable, and genuinely useful. A sought-after mentor and speaker, he connects industry, academia, and startups to turn emerging technology into practical impact.

How do you guarantee that the source code produced by an AI has high quality? And how do you measure it?

In times when everyone can build advanced software applications, discussions tend to focus on how to do this as efficiently as possible. Should you use integrations like MCP or CLI, orchestrations like subagents or agent protocols, or a prompting strategy like Spec Driven Development? Benchmarks like SWE-Bench measure how effective AI agents are at finding workable solutions, but very few seem to be talking about how to produce high-quality source code that not only runs optimally in terms of performance and memory usage, but is also well-documented and beautifully structured.

In this talk, Johan Sanneblad presents his production-proven process for creating complex applications with very high source code quality, achieving hundreds of thousands of downloads. You will learn how to create source code with AI agents that can be used as a software foundation for decades, instead of just focusing on how to produce code as quickly as possible. This process is already being used by some of the largest companies in Sweden, migrating old applications to modern tech stacks and creating new services in record time.

Johan Sanneblad, PhD, has worked with software development for large global organizations like Apple, EA, Google, Lego, Microsoft, Sony and Yamaha. Before starting TokenTek, Johan was Director of Human-centered AI at RISE. One of his most recent applications, Notebook Navigator for Obsidian, was the #1 most downloaded new plugin for the text editor Obsidian in 2025.

OpenEuroLLM is Europe’s answer to closed-source AI. With a massive strategic allocation of EuroHPC’s total capacity, including access to the upcoming JUPITER supercomputer, we are building the next generation of open European foundation models. In this talk, we unpack the reality of R&D for this continental operation. We will detail AI Sweden’s full-stack leadership: from defining scaling laws and custom tokenizers to curating data pipelines and driving the critical post-training phase. Supported by Vinnova, national and European partners, and integrated with other European projects such as TrustLLM and DeployAI, this project ensures digital sovereignty and green computing while retaining top-tier AI talent here in Sweden and across Europe.

Jonas Lind (PhD) is the Head of Research for the NLU team at AI Sweden. With over 25 years of experience in language technology, his career spans academia, AI for intelligence applications, speech technology, and forensic voice biometrics. Before leading NLU initiatives for OpenEuroLLM, Jonas worked extensively in national security for governmental agencies and as a founder of his own private company. He now oversees the post-training and refinement of large-scale European foundation models, ensuring they meet rigorous standards for performance, safety, and digital sovereignty.

Machine learning systems often fail in the most frustrating way possible: nothing crashes, training looks fine, and yet real-world performance is poor. Debugging these silent failures is one of the hardest and most underestimated challenges in applied ML. Most practitioners learn through painful experience, slowly building a personal “bag of tricks” from past mistakes and relying on intuition to guess what went wrong. In this talk, we explore how this intuition-driven approach emerges, why it sometimes works, and why it breaks down as ML systems grow more complex. We unpack how availability bias shapes where engineers look first, what they ignore, and why so many debugging efforts are often spent in the wrong parts of the pipeline. Finally, the talk introduces a structured approach to ML debugging. It starts from visible symptoms and uses small, targeted experiments to guide the next steps. This transforms debugging from guesswork into a systematic reduction of the search space, giving attendees a framework they can apply directly in their own ML projects.

Juliano is a Staff Engineer at Modulai, where he provides technical mentorship to machine learning engineers and develops ML systems across diverse industries using a range of ML techniques. He also founded Juliano Labs, an education business focused on high-quality workshops and courses on ML topics and structured problem-solving. Juliano brings a pragmatic and experiment-driven approach to ML debugging, shaped by 10 years of building and debugging ML pipelines across academia and industry.

No data should go unnoticed in the company. When we stumbled upon old, forgotten audio files gathering digital dust, our eyes lit up—these seemingly idle recordings turned out to be a rich source of insights, just waiting to be analyzed. We quickly mapped out the plan: download the files, transcribe them, extract the useful details—and just like that, a whole beautiful project was ready to roll! It may be hard to imagine all the little pitfalls, twists, and bumps that come with an initiative like this. How to measure conversation quality when all you have are mono-channel audio files? How to evaluate the AI system itself? How can prompts be updated safely without introducing risk? And how to make the solution fully compliant? We faced these challenges, and we’re excited to share our journey. The session will cover how we used AI to ensure quality and compliance in the contact center and the simple but effective solutions we applied.

Justyna Krok is a Data Lead at B2 Impact S.A. She began her journey in data and modelling by earning both a bachelor’s and a master’s degree in Informatics and Econometrics from AGH University of Krakow. Since then, she has worked in data science and machine learning roles, designing and delivering data-driven and AI-based solutions to address challenges in the debt management domain. To further develop her expertise, she recently started a PhD focused on multi-agent AI systems.

Implementing conversational BI in the private sector is hard; doing it for a public sector school administration is a different beast entirely. In December 2025, we kicked off a pilot with Göteborgs Stad. By the time we stand on this stage, we will have 3 months of real-world data, user feedback, and likely a few scars to show for it.

This talk moves beyond the hype of “chatting with data” to explore the practical reality of deploying AI in a municipality. We will share the unpolished truth of our journey: navigating strict compliance, managing expectations for a diverse, non-technical user base, and the technical realities of bridging modern AI and municipal data. A transparent look at what happened when we tried to bring AI to the classroom administration.

Karl Alfredsson is the Head of Development and Transformation at the City of Gothenburg’s Department of Education and a PhD candidate at University West. His work sits at the intersection of artificial intelligence, digital transformation, and modern public governance. Through real-world experimentation and action research, he explores how AI and data-driven systems are reshaping responsibility, trust, and learning in public organizations. With a background spanning education, leadership, and creative technology, Karl brings a forward-looking, human-centered perspective on how the public sector can evolve in the era of intelligent systems.

In many real-world vision systems, performance is limited less by model architecture than by the cost and quality of the data. This talk introduces a practical approach for "min-maxing" your data: maximizing expressibility while minimizing sample count through information-based subset selection. Whether choosing calibration frames or deciding which images deserve annotation, selecting the most representative samples can lead to faster results, lower labeling effort, and more accurate models, especially when building few-shot or data-efficient pipelines.

Taigatech is a Gothenburg-based startup that is making sawmills around the world more efficient with AI. Deployments are fast-paced, and AI models need to uphold quality quickly. In this context, smart data selection is not an academic luxury but a necessity for quick deliveries and maintainable neural networks.

Drawing on Taigatech's work in computer vision, I'll show how these techniques improve reliability and efficiency in production systems where data collection is expensive and conditions are harsh. Attendees will learn simple, deployable strategies for using their data more efficiently and building CV systems that perform better with less.

Mattias first encountered machine learning and computer vision in my bachelor's thesis when studying Engineering Physics at Chalmers. Fascinated by the subject, he continued by studying Complex Adaptive Systems, specializing in machine learning. 

Since then, he has worked in computer vision throughout his career, including serving as Taigatech's first hire and helping design multiple vision engines from the ground up. These products are used worldwide, improving yields and sustainability.

Beyond his professional work, he pursues several personal projects in machine learning, with a particular interest in reinforcement learning.

Unlabeled data is abundant in industry, but ground truth is scarce. Trafikverket captures hundreds of terabytes of high-resolution imagery annually to monitor Sweden’s railway network. The sheer volume of this data renders manual labeling—and therefore traditional supervised learning—unfeasible.

This talk explores how we are overcoming the "labeling bottleneck" by deploying a self-supervised foundation model trained on this massive, unlabeled archive. We will bridge the gap between theory and practice, starting with a comparison of Self-Supervised Learning (SSL) paradigms. Moving beyond theory, we will demonstrate the practical application of SSL in a large-scale industrial setting.

Attendees will learn how Trafikverket leverages SSL to learn robust visual representations, enabling the fine-tuning of models for critical downstream tasks—such as detecting rail cracks and identifying key infrastructure assets—with minimal labeled data.

Mladen Gibanica is a Senior Data Scientist at eghed, working primarily with Trafikverket to automate railway maintenance using computer vision. He was previously at Volvo Cars, initially as an industrial PhD student, and later as a data scientist working on projects from different domains.

Mladen holds a PhD in Applied Mechanics and has co-founded Ingenjörsarbete För Klimatet, a non-profit organisation conducting engineering work for a sustainable civilisation.

AI has transformed software development at unprecedented speed. AI-assisted development is now evolving into fully agentic systems that can orchestrate themselves and work independently. While autonomy is the goal, the missing link is grounding AI systems in high-fidelity, domain-specific context. Successful adoption, therefore, depends on leveraging MCP servers and specialized skills to anchor AI in the enterprise’s real-world environment.

This talk explores the technical and business foundations required to securely scale agentic capabilities. We will highlight the importance of clarifying intent through Specification-Driven Development, grounding AI models in domain context for reliable performance, and establishing a practical governance model to ensure the safe, compliant, and responsible integration of MCP servers.

Nasser Mohammadiha is an AI/ML expert specializing in the large-scale adoption of generative AI and enhancing R&D efficiency. With extensive experience spanning academia and industry, he is dedicated to fostering industry–academia collaboration and strengthening the Gothenburg AI ecosystem. Recently, his work has focused on high-impact R&D use cases, AI governance and risk mitigation, and building the internal communities necessary to scale generative AI across Ericsson.

Artificial Intelligence has become increasingly impactful in the life sciences over the last few years, pushing scientific boundaries, as exemplified by the recent success of AlphaFold2. In this talk, I will provide an overview of how AI has impacted drug design in the last few years, where we are now and what progress we can reasonably expect in the coming years. The presentation will focus on deep learning-based molecular de novo design; however, aspects of using multi-agent systems and the integration of AI with automation will also be covered.

Ola Engkvist, PhD, is Executive Director and Head of Molecular AI within Discovery Sciences at AstraZeneca R&D, where he leads the development and application of machine learning and artificial intelligence to accelerate drug design. He has published over 180 peer-reviewed scientific articles and is an ELLIS fellow at the European Laboratory for Learning and Intelligent Systems and a 2025 Clarivate Highly Cited Researcher. He holds an adjunct professorship in machine learning and AI for drug design at Chalmers University of Technology, serves as a Trustee of the Cambridge Crystallographic Data Centre, and is recognized for his work in pharmaceutical innovation.

Implementing conversational BI in the private sector is hard; doing it for a public sector school administration is a different beast entirely. In December 2025, we kicked off a pilot with Göteborgs Stad. By the time we stand on this stage, we will have 3 months of real-world data, user feedback, and likely a few scars to show for it.

This talk moves beyond the hype of “chatting with data” to explore the practical reality of deploying AI in a municipality. We will share the unpolished truth of our journey: navigating strict compliance, managing expectations for a diverse, non-technical user base, and the technical realities of bridging modern AI and municipal data. A transparent look at what happened when we tried to bring AI to the classroom administration.

P-A Gustafsson is the CEO and co-founder of Inblick.ai, an AI platform that turns organizational data into real-time, decision-ready insights. Today, his work focuses on applying AI in public-sector and educational contexts, helping school administrations and municipalities make better, faster decisions within strict governance and compliance frameworks. He began his career helping shape the classic video game Worms at Team17, joined DICE as one of its first designers, and later founded and exited the game studio Solidicon. Through his agency Future Memories, he has delivered 70+ projects for organizations including STC, Axfood and Mölndals Stad. P-A is also co-founder of Nestic.ai and an early investor and board member of health-tech scaleup Weekly Revolt.

A few years ago, building a machine learning system meant training your own model from scratch. Today, most teams start with an API call to a proprietary model, and that’s usually the right choice. But as products mature, the equation changes. Costs and control become limiting factors, fine-tuning becomes relevant, and open-source models are often considered.

In this talk, I’ll share what I’m seeing from working with AI startups and scaleups like Lovable, Suno, Reducto, and Cognition: when sticking with hosted APIs makes sense, when fine-tuning or switching to open source is the smarter move, and why some companies are still training their own models. We’ll look at the changing role of model training, and why it’s becoming rarer, but more important, than ever.

Rebecka Storm has a background in machine learning and has held data leadership roles at iZettle and Tink. After co-founding data orchestration startup Twirl, which was acquired by Modal, she now works on AI infrastructure, including serverless GPUs for model training and inference and sandboxes for executing AI-generated code. In 2018, Rebecka co-founded Women in Data Science Sweden, an organization that promotes inclusivity through conferences, mentorship programs, a speaker database, and other initiatives to inspire and support women working in data.

As the second-hand fashion industry grows rapidly, companies face increasing challenges in managing large product inventories—from identifying fashion items accurately to setting competitive prices across multiple resale channels. Bencha addresses these challenges with an automated pricing engine, real-time product identification, and a recommerce hub that guides clients on where, how, and at what price to sell their items. Under the hood, our systems process millions of unstructured product documents every day to power these capabilities.

For large-scale agentic and AI workflows, cost and service-level objectives (SLOs) such as latency and throughput quickly become bottlenecks when relying on large general-purpose models or proprietary APIs. At the same time, there is often headroom to improve accuracy and output quality for narrow, well-defined tasks. As model capabilities advance, one powerful pattern is to leverage large models for supervision while training much smaller specialized models that satisfy strict SLOs—with far lower variance, latency, and resource footprint—and still preserve quality or even outperform larger systems in domain-specific scenarios.

This talk focuses on how resource footprint, latency, throughput, and reliability constraints shape architectural and modeling choices in the context of open-weight AI. Demonstrated through a real-world production case study using Vision Language Models, it will detail Bencha’s systematic methodology for scoping an MVP, and iterating on data curation, fine-tuning, and evaluation strategies to reflect production behavior rather than benchmark scores. Attendees will gain practical strategies for extracting maximum value at scale—especially when defaulting to costly proprietary APIs is not an option.

Robert Cedergren is Founding Head of AI at Bencha, where he architects the company’s AI platform. He is a hands-on ML/AI engineer with a strong foundation in Computer Vision, NLP, and Generative AI, and has built and scaled production machine learning systems at multiple startups—spanning complex training workflows to engineering cloud-native runtimes that deliver both low latency and high throughput. Robert holds an MSc in Machine Learning from KTH and thrives in high-ownership roles shipping production-grade, end-to-end AI systems.

Soon, all AI workloads will demand the highest levels of performance and security—requirements that virtualized clouds increasingly struggle to meet. This talk introduces Bare Metal: physically isolated, purpose-built infrastructure granting direct access to raw compute power without layers of virtualization. We'll dissect this single-tenant architecture from A to Z, focusing on true data sovereignty through ownership and control of the entire supply chain end-to-end: data centers, GPUs, and networks. We'll also explore confidential compute and how to eliminate the leakage risks and "noisy neighbor" unpredictability common in shared environments. As a proof-of-concept, we'll showcase AI Sweden's "Svea" project, which leverages this sovereign infrastructure for public sector needs.

Co-founder and CEO of Airon, a Swedish technology company pioneering AI factories that transform Sweden's abundant electricity into high-value computing power. Robert launched the first facility in 2020, well before the AI boom, securing early partnerships with Nvidia and positioning Airon ahead of most European competitors. Airon maintains true sovereignty over the entire stack end to end: designing, building, owning and operating its data centers, owning the GPUs, and controlling the networking infrastructure.

In this talk, The Good AI Lab and Doctors Without Borders share how they are building a long-term collaboration where AI is taught and developed openly, accountably, and in service of humanitarian work. They will outline how AI enablement, strategy, and applied research projects can fit into a single roadmap, and what they have learned so far about making AI useful in high-stakes, low-resource settings without losing sight of dignity, context, and constraints.

Sofia Giannotti works in project design and innovation at Médecins Sans Frontières (MSF) Italy, where she develops cross-departmental initiatives to strengthen organizational learning, stakeholder engagement, and social impact. She develops participatory formats for volunteers and youth and supports the integration of digital tools and data-informed approaches into MSF’s strategies.

With a background in Politics, Philosophy and Economics and a Master’s in International Humanitarian Action (NOHA), Sofia combines human-centered design and analytical thinking to foster innovation within complex humanitarian systems. She is passionate about building inclusive processes that enable organizations to adapt and create meaningful change.

In this presentation, Sofia Tapani explores the transformative journey of Statisticians and Data Scientists in the age of AI and Automation. Drawing from her own evolution—from a hands-on coder to a strategic leader in drug development—Sofia illustrates how Statisticians are no longer just technical experts but central drivers of innovation, quality, and decision-making in modern medicine.

Key themes include:

• Strategic Evolution: How statisticians are stepping into leadership roles, shaping clinical strategy and regulatory pathways.
• Collaborative Intelligence: The power of interdisciplinary teams—biostatisticians, clinicians, programmers, data scientists—working together to solve complex challenges.
• AI Integration: Real-world examples of how AI and automation are streamlining clinical trials, enhancing analysis, and accelerating reporting.
• Data Storytelling: Strategies for translating complex data into compelling narratives that resonate across scientific and non-scientific audiences.
• Ethical AI Governance: Building robust frameworks to ensure responsible and safe use of AI in healthcare.
• Future Readiness: A call to action for Statisticians/Data Scientists to embrace continuous learning, innovation, and emerging technologies to shape the future of drug development.

This session will inspire attendees to rethink their role—not just as analysts, but as strategic catalysts in the AI-powered transformation of healthcare.

Sofia Tapani (PhD) is Executive Director of Statistics at AstraZeneca, leading strategic initiatives that advance the role of quantitative science in drug development. With a background in mathematics and biostatistics and a commitment to modernizing evidence generation, she champions the integration of advanced analytics and AI to improve decision‑making across clinical programs. Sofia is recognized for her contributions to industry thought leadership on the evolving role of statisticians in an AI‑driven future. She is passionate about fostering collaboration between statisticians, data scientists, and cross‑functional partners to accelerate impactful innovation.

In the depths of a Boliden mine, 700 meters underground, running traditional cloud-first AI on heavy-duty machinery falls short on delivering results. In a collaboration with Volvo Group and academic partners, we validated a novel approach: turning the trucks themselves into intelligent and interactively queryable computational databases. By embedding a tiny combined main-memory database and computation engine directly onto heavy-duty mining vehicles, we transformed the fleet into a distributed system where data streams are analyzed and queried at the source in real-time.

This talk shares the architectural lessons learned from deploying this "database-on-wheels" model to monitor critical metrics like battery health, energy regeneration, and driving patterns in a connectivity-constrained environment.

Beyond the immediate deployment, this architecture offers a fundamental shift in how we build industrial AI. We will explore how exposing physical assets via a familiar SQL-like interface paves the way for the next generation of Agentic Workflows. Instead of dealing with rigid firmware cycles, autonomous agents can simply “query” the fleet for insights or update model weights as easily as updating a row in a database table. We will discuss the trade-offs of edge-native computational query processing and how this approach decouples rapid AI innovation from the slower engineering cycles of heavy machinery.

Stefan Månsby is the CEO of Stream Analyze, an Edge AI company enabling intelligent decision-making on distributed devices. A serial co-founder and team builder, he co-founded Basefarm, growing it from a startup to a leading European managed services provider and has built innovation units, consulting practices, and AI teams throughout his 25+ year career. His experience spans CTO, CIO, and senior advisory roles at Orange Business and Basefarm, where he led projects in industrial IoT, fintech, and cybersecurity across three continents. And he still writes code.

This case concerns the development of an end-to-end geospatial pipeline that integrates curated spatial attributes with Husqvarna’s internal data to produce a coherent view of golf course environments. Developed in collaboration with Knowit Solutions, the pipeline establishes the technical foundation for geospatial analytics that drive business insights. Structured spatial indexing and robust metadata contribute to an annotated dataset that supports computer vision workflows in validating patterns and enriching geospatial analytics.

Husqvarna is further advancing its data-driven approach by using geospatial analytics to generate deeper insights into green spaces and understand regional variation. This evolving technical foundation is strengthening Husqvarna’s ability to combine visual and spatial information for more precise green space evaluation and identification, enabling more focused and informed decision-making across the organization.

Vivek Chaurasia is a full‑stack data scientist with a PhD in Computational Physics and extensive experience delivering AI solutions that drive measurable business value. Specializing in machine learning, computer vision, and large‑scale data engineering, Vivek has led end‑to‑end development of forecasting, segmentation, and prescriptive analytics systems used across global operations. Vivek builds scalable ML pipelines, strengthens data quality and governance, and translates complex analytics into actionable insights for commercial and technical stakeholders. With a strong track record of accelerating AI adoption, mentoring teammates, and shaping strategic data initiatives, Vivek consistently turns advanced analytics into tangible operational and financial impact.