Speakers

Update on the NVIDIA - TII Collaboration: First Wave of Results

This session provides an overview of ongoing collaborative work between NVIDIA and the Technology Innovation Institute (TII). Over the last couple of months we have initiated a number of joint collaborative projects including:

• Work aimed at improving the efficiency, performance, and deployability of TII’s language models. Our efforts focus on applying NVIDIA’s Puzzle methodology for model distillation and pruning to the Falcon series, with the goal of reducing computational cost and latency while preserving model quality.
• We investigate inference optimisation of Falcon-H1 using our latest quantisation technology including variants of FP8 and NVFP4. We are also working on integrating TII’s models into the NVIDIA Inference Microservices (NIM) framework, to enable streamlined deployment, scalable inference, and easier production adoption.
• In parallel, we are exploring Post-NAS optimization techniques for Mixture of Experts (MoE) models using NVIDIA’s Jet-Nemotron methodology, enabling efficient exploration and optimization of attention block designs in MoEs.
• Complementary to these efforts, we are benchmarking and optimizing TII’s models on NVIDIA GPUs, and applying model quantization to further enhance inference efficiency.
• Finally, we provide a high level update regarding our physical AI collaboration.

The goal of the talk is to share with the audience the first wave of results as well as trajectory for further collaboration.

Dr. Adam Grzywaczewski is a senior deep learning data scientist at NVIDIA, where his primary responsibility is to support customers in scaling and deployment of their deep learning workloads. Adam also leads NVIDIA’s EMEA AI SA organisation. Over more than a decade, Adam has specialized in large-scale DNN training, focusing not just on deep learning system and software design, but also on algorithms that allow for large batch data/model/pipeline parallel training. Adam works with customers with high computational needs, including key automotive customers and organizations with the need for large scale NLP and conversational AI. He also leads an EMEA AI team that partners with NVIDIA’s strategic customers to address the challenges of building and deploying deep learning models, while serving as a key interface to NVIDIA Engineering, Product, and Research groups. Adam also focuses on techniques for computationally efficient inference, including parameter/compute efficient model design and model pruning and quantization. Prior to joining NVIDIA, Adam was responsible for building up the UK government’s machine-learning capabilities while at Capgemini, and worked in the Jaguar Land Rover Research Centre, where he worked on the self-learning car project.

The Triple Helix of Cultural Fit in LLMs: The Aim, The Facts, and The Complex Reality

Large language models inherit cultural biases from training data dominated by English and Western perspectives. While techniques like data rebalancing and safety alignment attempt to address these biases, cultural nuances extend beyond language representation—it permeates values, norms, and seemingly neutral texts that reflect specific worldviews. This presentation examines cultural alignment in LLMs through a triple helix framework: developer aspirations for inclusive systems, observations captured in biased data, and the realization of these biases in model outputs. Through concrete examples, we demonstrate how ostensibly neutral content carries implicit cultural assumptions. We advocate for proactive engagement with training data and creative use of existing capabilities to identify and mitigate cultural misalignment, developing more equitable systems for diverse global communities.

Dr. Ahmed Abdelali is a Senior Principal AI Researcher specializing in natural language processing and machine learning, with emphasis on Arabic language applications. With over 25 years of experience across academia and industry, he possesses advanced expertise in deep learning, computational linguistics, and multilingual AI systems. He received his PhD in Computer Science from the New Mexico Institute of Mining and Technology. Before joining Humain, he worked at the National Center for AI, Saudi Authority for Data and AI, and previously spent 10 years at Qatar Computing Research Institute and 13 years at New Mexico State University. Dr. Abdelali has led and contributed to the development of prominent projects including ALLaM, Farasa, Shaheen, and NatiQ, as well as NSF- and DARPA-funded initiatives. He has published over 100 research papers in peer-reviewed conferences and journals.

Reimagining Enterprise Services Through Language Models, Agent Automation Middleware, and Quality Control AI

AI-enabled agents are redefining enterprise and service operations through reasoning, autonomy, and self-optimization. The emerging wave of agentic AI systems, powered by large language and action models, promises to extend automation beyond static workflows toward adaptive, goal-driven intelligence. These systems can perceive, reason, and act across domains such as customer service, healthcare, banking, logistics, and telecom, dynamically optimizing performance, integrating with tools and services, and continuously improving through feedback. As these capabilities expand, robust quality control becomes a foundational requirement, ensuring reliability, trustworthiness, compliance, and operational consistency. Quality Control AI including grounded validation, automated safeguards, and feedback-driven correction enables agentic systems to operate safely and predictably at scale. This talk revisits the abstractions and middleware needed for AI-powered services, focusing on the convergence of service integration, agent self-improvement, multi-agent orchestration, and built-in quality control as foundations for the next generation of adaptive, robust, interoperable, and intelligent enterprise services.

Prof. Boualem Benatallah is a full professor of computing at Dublin City University (DCU, Ireland) since January 2022. He has had over 21 years as a research leader and academic at UNSW Sydney (Australia), where he served as senior lecturer, associate professor, professor, and Scientia Professor, before joining DCU. He is a Fellow of the IEEE.

His main research interests are in AI-enabled services, process AI, LLM-powered agents, quality control in crowdsourcing and AI services, service-oriented computing, and business process management. He has published more than 330 refereed papers, including more than 90 journal papers. Most of his papers appeared in very selective and reputable conferences and journals. His research attracted a large amount of competitive research funding through national and international grants from both government and industry. He supervised over 38 research students to completion. He was awarded the prestigious IEEE TCSVC Research Innovation Award for contributions to Model-driven Web Services Composition. He has also won multiple best paper awards at prestigious conferences and received the IBM Faculty Award. With his co-authors, he was recognized by IEEE TSE for one of the most influential papers of the journal’s 3rd decade and contributed a retrospective to its 50th anniversary issue in 2025.

Boualem has been general and PC chair of a number of international conferences and has served as guest editor of several special issues in leading journals. He was a member of the Steering Committee (SC) of the BPM conference. He is a member of the SC of the ICSOC and CoopIS conferences and serves on the editorial boards of several prestigious journals including ACM Transactions on the Web, IEEE Transactions on Services Computing, and ACM Computing Surveys. He was a member of the team (comprising multiple university, government, and industry partners) that founded and constructed the successful bid for the Smart Services CRC (Cooperative Research Centre, Australia). He was also research leader of the Data Curation Foundry research stream at the Data to Decisions CRC (Australia). He is funded investigator at the Insight Research Centre (Ireland).

Apertus: The Transparent & Sovereign Open Source Swiss LLM

Apertus represents a paradigm shift in open-source language model development, offering a fully transparent, 70-billion-parameter multilingual model trained on 15 trillion tokens across 1,000+ languages with 40% non-English data, making it the first large-scale LLM to prioritize underrepresented linguistic communities. Developed collaboratively by Swiss institutions (EPFL, ETH Zurich, and Swisscom as initial strategic partner) and trained on the Alps supercomputer with sustainable infrastructure, Apertus provides complete reproducibility through open weights, training recipes, and full documentation while adhering to Swiss data protection and EU AI Act compliance requirements. This work establishes a blueprint for trustworthy, sovereign AI development, demonstrating how public institutions can build inclusive foundational models that serve both research and commercial applications while maintaining ethical standards and respecting data ownership through machine-readable opt-out mechanisms.

Dr Daniel Dobos is Research Director at Swisscom. He is responsible for relations with universities, universities of applied sciences and other research institutions. Together with employees from all Swisscom business areas, he and his team develop solutions that use the latest research and technology developments for the benefit of Swisscom customers. Previously, he led research and AI data analysis projects at the CERN research centre and at the United Nations.

The Salamandra Family: Advancing Open Foundation Models through Pre-Training and Post-Training recipes.

This talk presents the Salamandra family of open foundation models developed at the Barcelona Supercomputing Center (BSC), focusing on the engineering and research insights behind their design. We will discuss large-scale pre-training on HPC infrastructure, efficient fine-tuning and post-training methods for long-context reasoning, and ongoing work on multimodal and multilingual enhancements. The talk will also cover lessons learned in scaling distributed training, optimizing inference, and ensuring reproducibility in open-source environments. By sharing both successes and challenges, the goal is to contribute to a broader community effort toward transparent and efficient large-model development.

Daniel Tamayo Mela is a Research Engineer in Natural Language Processing at the Barcelona Supercomputing Center. He has worked on the development and large-scale pre-training of both encoder and decoder language models, including contributions to the Salamandra model family and to interpretability research. His work focuses on high-performance training infrastructure, knowledge distillation, and efficient model deployment, with an emphasis on improving the efficiency and scalability of LLM training pipelines.

Why do we need a sovereign infrastructure to run Open Source LLMs?

A sovereign infrastructure is crucial for running open source large language models (LLMs) because it empowers organizations and nations with control, transparency, and security over AI systems and the sensitive data they use. True sovereignty in AI comes from owning and managing the entire technological stack and data pipeline, ensuring autonomy from external vendors or cloud providers.

This talk will demonstrate how building and operating open source LLMs on sovereign infrastructure enables protection of intellectual property, compliance with local data laws, and the creation of customizable, stable platforms to fuel innovation and meet strategic needs. Such infrastructure ensures organizations and nations can avoid geopolitical risks, enhance data privacy, and align AI development with local priorities, values, and regulations. Sovereign infrastructure ultimately supports economic growth, national security, and the responsible use of AI at a local scale.

David Gurle is a French engineer, a serial entrepreneur and pioneer in IP communications and distributed computing. He is the founder and former CEO of Symphony Communication Services, a $1.5 billion unicorn providing secure messaging for financial institutions, and previously led business units at Skype, Microsoft, and Thomson Reuters.

Currently, David is the founder and CEO of Hivenet and PoliCloud, building distributed cloud software and data center infrastructure. He holds a master’s degree in computer science and telecommunications from EFREI, and his career is marked by launching innovative platforms and advancing secure communications and cloud technologies globally. David Gurle lives in Dubai, UAE.

Rethinking the “evaluation” of AI—powered systems

Abstract: AI-powered services follow a very different dev pattern than traditional ones. Very often, progress is measured as improvement in some quality score rather than new features. This is somewhat similar to what Karpathy calls "software 3.0", as opposed to software 1.0 ("traditional" sw dev) and 2.0 (ML model dev). This is not very different to how we train AI models: we iterate on a system, measure the quality, identify where to improve, and try to make progress by moving in the right direction. Many "human-powered" services behave in the same way. If measures of quality drive improvement cycles, then it is clear that such measures need to be "accurate". In fact, many argue that if we have a good eval, then AI itself can take care of improving the AI-powered service. On the other hand, the more noisy are the measures, the harder it will be for dev cycles to improve, if at all. The "eval" process is therefore foundational in AI services. Software 2.0 teams are aware of this and eval there is relatively more mature than in other software teams, where quality is strongly influenced by "traditional" testing models and methods from software 1.0. The harsh reality is that while eval is foundational and while AI makes the dev process "easier", the eval process is actually really hard. On top of that, and perhaps most importantly, teams without an AI-first culture tend to underestimate how hard that is. This talk will describe how brittle the "eval" process is for AI services and how we can make it more reliable, thereby dramatically accelerating the speed and efficiency at which our services improve. We will see the pitfalls we need to avoid and the tricks we can use to make our "eval" stronger. As we will see, the main "trick" is giving the right name to what we do. Per se, in this talk you will learn no new knowledge: I will simply stitch together many "obvious" concepts to paint what I hope is a useful picture of how to guide "eval" processes.

Fabio Casati is Professor at the University of Trento and principal AI architect for ServiceNow. Fabio focuses on designing, architecting and deploying AI-powered workflows for enterprise customers. On the research side, he is working on evaluations and governance of AI systems and on AI systems that serve needs of individuals and subjective point of views.

Toward foundation models for neuro-technologies,: real-time neural decoding with hybrid state-space models

Much like large language models have done for text, neuro-foundation models aim to capture general and universal patterns of neural activity from different acquisition modalities, across brain regions, tasks, and even species. The learned representations can then be used to fine tune for varied tasks from diagnostic assistance, to machine learning tools for neuro-technology. I this talk, I will focus on real-time decoding of neural activity. This is central to neuroscience and neuro-technology applications, such as closed-loop experiments and brain-computer interfaces where models are subject to strict latency constraints. Traditional methods, including simple recurrent neural networks, are fast and lightweight but often struggle to generalize to unseen data. In contrast, recent Transformer-based approaches leverage large-scale pre-training for strong generalization performance, but typically have much larger computational requirements and are not always suitable for low-resource or real-time settings. To address these shortcomings, we present POSSM, a novel hybrid architecture that combines individual spike tokenization via a cross-attention module with a recurrent state-space model (SSM) backbone to enable (1) fast and causal online prediction on neural activity and (2) efficient generalization to new sessions, individuals, and tasks through multi-dataset pre-training. I will present of series of results showcasing the efficiency of our approach, ranging from rapid low-latency inference of pre-trained models to transfer learning between species.

These results suggest that hybrid SSMs are a promising approach to bridging the gap between accuracy, inference speed and generalization when training time-series models for real-time closed-loop applications

Guillaume Lajoie is an Associate Professor in the Department of Mathematics and Statistics at Université de Montréal and a Core Academic Member of Mila – Quebec Artificial Intelligence Institute. He holds a Canada CIFAR AI Research Chair, and a Canada Research Chair in Neural Computation and Interfacing. His research is positioned at the intersection of AI and Neuroscience where he develops tools to better understand mechanisms of intelligence common to both biological and artificial systems. His research group's contributions range from advances in multi-scale learning paradigms for large artificial systems, to applications in neurotechnology. Dr. Lajoie is actively involved in responsible AI development efforts, seeking to identify guidelines and best practices for use of AI in research and beyond.

Transforming Metabolomics Through Machine Learning and Generative AI: From Automated Workflows to Precision Medicine Applications

The convergence of machine learning, generative AI, and metabolomics is revolutionizing our approach to biomedical research and clinical applications. This presentation will explore cutting-edge innovations that are accelerating discovery and translation in life sciences.

I will introduce GigaKit and GigaView, our integrated platforms for high-throughput metabolomics data processing and visualization, demonstrating how these tools enable researchers to handle massive datasets with unprecedented efficiency. The talk will showcase how agentic AI systems are transforming life sciences workflows, moving from passive analytical tools to active research partners that can design experiments, interpret results, and generate hypotheses.

A key focus will be on Lab-in-a-loop architectures, where AI systems continuously learn from experimental outcomes to optimize protocols and accelerate discovery cycles. I will present case studies demonstrating the integration of organ-on-a-chip technologies with AI-driven metabolomics, creating powerful models for drug discovery and personalized medicine.

Key takeaways will include practical strategies for implementing AI in metabolomics laboratories, overcoming common challenges in data integration, and future directions for the field as we move toward truly personalized, AI-driven medicine.

Dr. Jacques Corbeil is a pioneering researcher operating at the intersection of machine learning and omics sciences, with over two decades of experience in medical genomics and bioinformatics. As the former Canada Research Chair in Medical Genomics (Tier 1, 2003-2024), Dr. Corbeil has established himself as a leader in applying artificial intelligence to complex biological challenges.

His research leverages state-of-the-art computational approaches to transform the way we diagnose diseases, predict treatment outcomes, and understand biological systems. Dr. Corbeil specializes in developing novel machine learning algorithms for interpreting the massive datasets generated by modern genomics and metabolomics platforms, including high-throughput mass spectrometry and next-generation sequencing.

His laboratory investigates critical biomedical questions, including host-pathogen interactions, the impact of antibiotics on microbial communities and environmental health, and the design of targeted therapeutics for infectious diseases and cancer. A particular strength is his expertise in integrating multi-omics data, combining metabolomics, genomics, and clinical data to create comprehensive disease models.

Dr. Corbeil maintains extensive collaborations with industry partners, helping organizations implement AI strategies and optimize their analytical processes. His work bridges the gap between computational innovation and clinical application, with a focus on translating big data analytics into actionable insights for precision medicine. His contributions have been instrumental in advancing our understanding of infectious disease dynamics and cancer progression through the lens of systems biology and artificial intelligence.

Using Tools and Agents to Overcome Model Boundaries

Large Language Models (LLMs) have demonstrated remarkable abilities across a wide range of tasks—from reasoning and summarization to coding and creative writing. Yet, despite their versatility, LLMs also exhibit clear limitations: they hallucinate facts, struggle with precise computation, and lack persistent memory or real-world awareness. This talk explores where and why LLMs fail, and how agentic systems and tool use can fill these gaps. We’ll look at the emerging ecosystem of techniques for extending LLM capabilities, specifically tool calling. The discussion will focus on practical pre and post training and prompting considerations for effectively interfacing with external functions. By the end, you’ll have a clearer understanding of how to equip LLMs with tool calling, choose the right tool calling strategies, and design hybrid systems that turn static models into dynamic problem solvers, particularly for Arabic centric applications.

Dr. Kareem Darwish is a principal scientist at the Qatar Computing Research Institute (QCRI), and he is co-leading the effort to create Fanar, an Arabic LLM. Previously, he was a principal scientist at aiXplain Inc working on efficient human-in-the-loop ML and speech processing. He was also the acting research director of the Arabic Language Technologies group (ALT) at QCRI where he worked on information retrieval, computational social science, and natural language processing. Kareem Darwish worked as a researcher at Microsoft and IBM and taught at the German University in Cairo and Cairo University.

His research on natural language processing has led to state-of-the-art tools for Arabic processing that perform several tasks such as part-of-speech tagging, named entity recognition, automatic diacritic recovery, sentiment analysis, and parsing. His work on social computing focused on predictive stance detection to predict how users feel about an issue now or perhaps in the future, and on detecting malicious behavior on social media platform, particularly propaganda accounts. His innovative work on social computing has received much media coverage from international news outlets such as CNN, Newsweek, Washington Post, the Mirror, and many others.

Aside from the many research papers that he authored, he also authored books in both English and Arabic on a variety of subjects including Arabic processing, politics, and social psychology.

AgentScope: A Developer-Centric Framework for Building Scalable Multi-Agent AI Systems

AgentScope is an open-source, Python-based framework designed to simplify the development, orchestration, and deployment of LLM-powered multi-agent applications. Built with a modular and asynchronous architecture, AgentScope enables developers to create complex agentic workflows with high flexibility, reliability, and safety—including support for tool-based interactions, real-time communication, and production-grade sandboxed execution . In this talk, I will demonstrate how AgentScope lowers the barrier to building real-world multi-agent systems, showcase practical use cases, and discuss its role in advancing open-source generative AI ecosystems.

Dr. Farruh is an expert in generative AI with extensive experience combining theory with practice. He has participated in various AI-linked projects and researched foundation models. He has a notable track record with over ten technical/scientific publications.

Why Agentic AI Is Moving Fastest in the Open

Agentic AI is evolving faster in open ecosystems than in closed ones. This session presents AWS's perspective on why open-source models, protocols, and frameworks are accelerating the entire field—and what that means for builders. We'll examine the full stack: how open models enable rapid customization and domain adaptation, how protocols like MCP and A2A are creating interoperability across vendors and clouds, and how open SDKs like Strands are making complex orchestration patterns accessible to more developers. Drawing from the AWS MCP Servers project (6.9k+ stars, millions of requests) and production deployments across industries, we'll share what we're learning about the architectural patterns emerging in the community and why supporting multiple open frameworks matters more than promoting any single approach. The thesis is straightforward: the highest velocity innovation in agentic AI is happening in the open, and the platforms that enable that openness will define the next decade of AI applications.

Laith Al-Saadoon is a Principal AI Engineer at Amazon Web Services (AWS), where he works on open-source agentic AI tooling and customer prototypes. He is the creator of AWS MCP Servers, an open-source project with 6.9k+ stars that has become a de facto standard for AWS MCP implementations, serving millions of monthly requests. He is frequently consulted by AWS product teams on agentic AI strategy and emerging AI capabilities.

With nine years at AWS, Laith has architected AI solutions for Fortune 500 enterprises including United Airlines, AutoNation, and PagerDuty. His work spans agentic AI systems, real-time voice AI, and generative AI applications that have been featured in AWS keynotes by Dr. Werner Vogels and Matt Garman, AWS CEO. He has authored eight AWS blog posts on machine learning, generative AI, and data analytics.

Beyond AWS, Laith contributes to open-source AI frameworks including LangChain, Strands Agents SDK, and Mem0. He co-authored "Minding the Machines," a whitepaper on AI ethics in criminal justice systems with the Center for Justice Innovation.

Laith holds a B.S. in Biomedical Science from La Sierra University, where he built computational genomics pipelines for DNA sequencing and conducted genome-wide association studies linking single nucleotide polymorphisms to body size phenotypes—an interdisciplinary foundation in large-scale data analysis and statistical inference that shaped his transition into AI engineering.

Falcon-H1: A Family of Hybrid-Head Language Models Redefining Efficiency and Performance

In this talk, we present the Falcon-H1 series of large language models, ranging from 0.5B to 34B parameters. Falcon-H1 series introduces an innovative parallel–hybrid attention–state space (SSM) architecture, which combines the performance and precision of attention with the efficiency of SSM. The largest Falcon-H1-34B model rivals the performance of leading open models, such as Qwen-3-32B, upon release. The smallest Falcon-H1-0.5B model remains the top-performing model to date. In the talk, we will walk through the rigorous methodology behind building the Falcon-H1 models and highlight the key innovations. We cover (i) architectural design, including the experimentation behind the parallel–hybrid design with adjustable SSM/attention ratios and associated mixer parallelism for improved efficiency; (ii) training dynamics aspects, from stabilizing large learning rate training of mamba-based model to custom maximal update parametrization (μP); and (iii) data-related insights, including multi-epoch training and anti-curriculum data strategies.

Maksim Velikanov is a Senior Researcher at the Technology Innovation Institute (TII) in Abu Dhabi, where he is part of the Falcon Foundational Models team and has contributed to several generations of Falcon models, including Falcon-Mamba and Falcon-H1. His main focus is on large language model training dynamics, while also being involved in other aspects of the LLM lifecycle from architecture design to data strategy. Concurrently, Maksim is completing a PhD in Applied Mathematics at École Polytechnique, working on optimization and generalization of quadratic problems with power-law spectrum, for which he received Yandex Ilya Segalovich Award for young researchers. His background is in physics, with MSc and BSc work on condensed matter theory, and a gold medal at the 2013 International Physics Olympiad.

AI, Open Source, and the SDGs: Building a System of Systems for Global Good

This talk examines how open and interoperable AI systems can operationalize the Sustainable Development Goals (SDGs) through a dual technical and governance lens. At the meta level, it focuses on using AI to connect the different SDGs targets and redesigning the SDGs framework as a system of systems. The goal is to reveal hidden patterns, dependencies, and interactions across goals, enabling more integrated and data-driven decision-making. By linking sectoral AI models and datasets through open standards, modular APIs, and transparent governance frameworks, this approach aims to turn the SDGs into an interconnected ecosystem rather than a collection of isolated targets. Such integration enables interoperability, collective intelligence, and real-time monitoring of progress across SDGs.

At the SDG implementation level, the talk presents concrete examples of AI applied to global goods in health, agriculture, education, and climate resilience, demonstrating how open-source approaches lower costs, enhance reproducibility, and improve trust. The presentation also discusses the need for common benchmarks, lifecycle energy efficiency metrics, and context-aware evaluation frameworks to ensure that AI for the SDGs remains accountable, sustainable, and globally accessible.

Dr. Mehdi Snène is Senior Advisor on Artificial Intelligence and Digital Transformation to the United Nations Secretary-General’s Envoy on Technology, where he also serves as Head of the AI and Digital Program. He leads global initiatives on AI governance, capacity building, digital public infrastructure, and chairs the United Nations Open Source Conference. His work brings together governments, international organizations, and research institutions to advance the responsible and inclusive use of AI. Before joining the UN, Dr. Snène was the Scientific and Technical Director of the Human Brain Project at EPFL, leading multidisciplinary research at the intersection of neuroscience, neuromorphic, HPC and artificial intelligence. His background spans AI policy, ethical frameworks, and applied research in health, neuroscience, and complex systems.

Mohamed El Amine Seddik is a Principal Researcher at the AI Research Centre of the Technology Innovation Institute (TII) in Abu Dhabi, where he leads the Post-Training & Reasoning Team. His research focuses on advancing the post-training and reasoning capabilities of large language models through innovative algorithmic and theoretical approaches. Before joining TII, he was a Machine Learning Researcher at the Huawei Research Center in Paris, where he worked on applied AI research. He previously held a postdoctoral position at École Polytechnique, conducting research at the intersection of graph neural networks and large language models. He obtained his Ph.D. from CentraleSupélec, Université Paris-Saclay, where his work focused on random matrix theory and its applications to modern machine learning. His research interests span foundational machine learning, random matrix theory, and both the theoretical and practical frontiers of large language models.

How Spotify builds with and for the open source community

At Spotify, openness is central to how we work. We contribute to the open-source ecosystem through research, publications, code, and platforms such as Backstage, while collaborating closely with the community that helps these projects thrive. As AI becomes increasingly integrated into R&D, our engagement with open source grows even more relevant, shaping how we build, share, and learn together. This talk will highlight how openness, community collaboration, and AI are coming together at Spotify to advance R&D experience and innovation.

Mounia Lalmas is Senior Director of Research at Spotify and Head of Tech Research in Personalisation, advancing state-of-the-art in personalization. Her work spans search, recommender systems, user engagement, and evaluation, with a growing focus on how Generative AI is transforming large-scale retrieval and recommendation. She also holds an honorary professorship at University College London and is a Distinguished Research Fellow at the University of Amsterdam. Previously, she was Director of Research at Yahoo and Professor of Information Retrieval at Queen Mary University of London. An active member of the search and recommender systems communities, she has co-chaired major conferences including SIGIR, WWW, WSDM, and CIKM, and authored over 260 papers. She was nominated for the VentureBeat Women in AI Awards for Research in 2022 and 2023.

MaxText: Achieving Simplicity and Top-Tier Performance for LLM Training with JAX

Training and scaling Large Language Models (Dense LLMs or MoEs) traditionally involves significant engineering complexity, often requiring specialized, low-level optimizations for specific hardware. This presentation introduces MaxText, an open-source, high-performance LLM framework built on JAX. We demonstrate how MaxText serves as a scalable and simple reference implementation for training on Google TPUs and GPUs. By leveraging the power of the JAX and XLA compilers, MaxText achieves state-of-the-art Model FLOPs Utilization (MFU) while remaining easy to modify. We will cover MaxText's architecture and its capabilities for both pre-training and post-training (including Supervised Fine-Tuning and Reinforcement Learning). Attendees will learn how MaxText provides a robust, forkable foundation for ambitious research and production projects, supporting a wide range of models like Llama, Gemma, and Mistral, DeepSeek out of the box.

Newfel Harrat is an engineering director at Google Cloud, where he oversees the Open Source Software AI/ML Software Stack. In this role, he focuses on AI/ML Frameworks designed to accelerate both model training and inference. Prior to his current position, Harrat led the Gemini Code Assist team at Google, where he was instrumental in developing AI-powered code generation tools that significantly enhanced developer productivity.

Newfel is dedicated to utilizing AI to enable developers and streamline their product development. He is also a fervent supporter of diversity and inclusion within the technology sector.

Newfel's career before Google Cloud included leading software development for Google Chrome OS.

Before joining Google, he led an engineering team focused on AI-driven incident prevention software for automotive and large enterprise fleets.

Newfel also held senior leadership positions at several Fortune 100 companies (Intel, Qualcomm, ADI), where he oversaw global software development teams.

He earned both a Bachelor of Science and a Master of Science degree in Electrical Engineering from the University of Massachusetts, specializing in Neural Network and Stochastic Signal Processing.

Open-Source Data & Tools for AI Model Development

Open-source large language models have experienced explosive growth over the last few years, matching and often surpassing the performance of closed-source Frontier AI models on various downstream tasks. Underlying this growth is the ecosystem that has rapidly evolved around open-source data and training software stacks, covering all phases of AI model development, ranging from real & synthetic data preparation to model training, fine-tuning, and benchmarking. In this talk, we will describe a suite of data preparation tools and recipes that IBM Research has developed for the training of the Granite model series and contributed to the open-source community.

Petros Zerfos, Ph.D., is Principal Research Scientist & Manager at IBM Research, Yorktown Heights, NY. He is the chief architect of the large-scale data engineering for IBM’s Granite AI Models and the platform for Generative AI model development and customization. He received his PhD and MSc in Computer Science from the University of California, Los Angeles (UCLA), CA, USA, and his M.Eng. in Electrical & Computer Engineering from the National Technical University of Athens, Greece. He has (co-) authored 65+ scientific papers published in top-tier journals & technical conferences in the areas of artificial intelligence, Big Data platforms, time series analysis, cloud and systems management, and has been designated IBM Master Inventor, having filed 60+ patents worldwide with 45 granted to date, while technologies that he has developed with his team have been productized in 20+ IBM products and services. Petros is also serving as technical advisor to the “Pharos” Greek AI Factory strategic initiative, providing strategic and technical guidance on the development of the EU AI Factory in Greece.

EuroLLM: A Family of Encoders and Decoders for Europe

In this talk, I will present EuroLLM, our family of multimodal encoder and decoder models developed to strengthen Europe’s multilingual and culturally aligned AI ecosystem. By emphasizing linguistic diversity, open collaboration, and data sovereignty, EuroLLM aims to advance natural language understanding and generation across European languages and domains.

I am an Associate Professor at CentraleSupélec and Co-founder / Chief Science Officer at Equall.ai, a legaltech startup where we developed the first LLM specialized for law and AI-driven workflows, enabling lawyers to work faster and more efficiently. My research focuses on Natural Language Processing and Large Language Models (LLMs), including large-scale training (EuroBERT, EuroLLM, SaulLM, CroissantLLM, TowerLLM), LLM efficiency and adaptation, and the development of novel evaluation metrics recognized at leading venues (ACL, EMNLP, NeurIPS, ICML, AAAI).