Opening Keynotes

Most AI agents fail in regulated environments because they lack the right architecture. This technical session reveals what you actually need: deterministic orchestration for auditability, explainable reasoning for compliance, governed tooling for security, and sovereign infrastructure for data residency.

Through a production insurance claims system processing thousands of decisions monthly, see how these components integrate into trustworthy, auditable systems. Learn practical lessons on confidence calibration, audit log management, and gradual autonomy scaling.

Walk away with architecture patterns and production checklists for deploying agents in finance, healthcare, insurance, and government.

Master Stage

Agentic Stack for Regulated Industries: Architecture Essentials

Agents are starting to handle real tasks.

They write and refactor code. They trigger tools. They take decisions inside workflows that used to be fully manual.

This panel looks at what it takes to make these systems reliable in production.

We explore how agents plan, act, recover from mistakes, and integrate with engineering teams.

A clear and practical look at how agentic AI is built today and how it is reshaping the way software is created.

Agents that actually do the work: how autonomy changes the way we build

In this talk, Bart Butler, CTO of Proton, will explore how Proton brings artificial‑intelligence capabilities to its privacy‑focused ecosystem while staying true to its zero‑access encryption principles. Drawing on the development of Lumo, the company’s conversational AI assistant, Bart will walk through the real‑world compromises and engineering decisions that make privacy‑preserving AI feasible at scale. Expect a deep dive into the challenges of combining security, scalability, and intelligence and how AI can be both powerful and privacy-first.

AI and Privacy: Building secure intelligence at scale

AI systems are evolving at breakneck speed: new architectures, model weights, and deployment frameworks appear weekly. But which innovations actually move the needle for developers?

Artificial Analysis will go beyond leaderboard hype to unpack the real technical trade-offs shaping today’s AI frontier: reasoning efficiency vs. latency, cost vs. intelligence, open vs. closed weights, and throughput vs. accuracy.

Drawing on months of live benchmarking across models, frameworks, and chips, they’ll share data-driven insights on how these forces interact and what it means for anyone building or scaling AI-powered systems in production.

Expect benchmarks, performance graphs, and a candid look at where the frontier actually stands.

Benchmarking the frontier: what really matters in the new AI stack

As AI models scale, so does their hidden cost: CO₂.

While few teams have the tools or methodology to accurately measure the carbon impact of inference, new frameworks are emerging to make energy use visible and actionable.

In this session, Salesforce and Scaleway will share how they’re building the foundations for transparent and standardized measurement of AI’s environmental footprint:

Salesforce will present the AI Energy Score, a standardized method developed with Hugging Face to evaluate model energy efficiency.

Scaleway will show how it translates hardware and GPU utilization data into carbon metrics through its Environmental Footprint Calculator.

A practical deep dive for ML and infra engineers who want to make “energy-efficient AI” measurable, not mythical.

Beyond FLOPs: Engineering transparency into AI’s carbon footprint

AI is growing fast and pushing every part of the stack. In this talk, Sean Varley from Ampere Computing explores how efficient compute shapes the next generation of AI infrastructure.

He looks at the hardware and software choices that let clouds run smarter models, faster iteration cycles, and AI applications that spread everywhere.

From inference economics to energy efficiency, this session highlights the practical decisions that help teams keep up with the pace of AI without exploding cost or power budgets.

Building Al that scales: efficient compute for a smarter, faster, everywhere world 

Agentic AI pushes infrastructure to its limits. To train models that see, understand, and act, teams need high-throughput storage, clean pipelines, and architectures that scale with evolving datasets.

H Company and VAST Data unpack how they approach these challenges: designing agent-oriented models, managing fast-moving training workflows, and building the data backbone that keeps iteration cycles short.

A focused look at the engineering trade-offs behind the next generation of AI systems.

Building at the speed of agents: from training pipelines to data infrastructure

How do we turn powerful AI models into robots that can reliably act in the physical world? Most robots still live in structured, predictable environments but the next frontier is the messy, human world: homes, hospitals, and care settings.

This session explores the technical leap from perception to action, and what it takes to deploy robots among us:

A deep dive into the next generation of assistive and expressive robots, built not just to operate, but to live alongside people.

From Foundation Models to Real-World Actions

Europe’s AI landscape is evolving fast, from open research labs to production-grade language models.

This session brings together Kyutai, the lab behind open research models like Moshi, and Indigo.ai, developers of Italy’s DistilBERTino and advanced conversational systems.

Together, they’ll discuss how European teams are translating cutting-edge research into usable, scalable AI products.

From lab to product with European voice model

As AI agents move from experimentation to production, the real challenge isn't building a single agent BUT orchestrating fleets of specialized agents that collaborate with each other and humans. Dust, is building the infrastructure layer for this new paradigm.

In this talk, you'll explore the practical architecture decisions behind production agentic systems: how to pragmatically combine RAG, MCP, and context windows; manage permissions across high-stakes and low-risk workflows; design multi-agent architectures where agents coordinate to solve complex tasks; and tackle the thorny problem of evaluation when traditional benchmarks don't capture operational reality. 

Discover Dust's learnings about enabling agent-to-agent-to-human collaboration at scale.

From single agents to agent fleets: lessons from building Dust's agentic platform

Steeve Morin of ZML will walk through how to build inference systems that run efficiently across a variety of hardware — from edge devices to servers — without sacrificing latency, throughput, or reliability. He will dive into architectural decisions, hardware constraints, trade-offs between generality and specialization, and strategies to make inference flexible, portable, and performant. Expect real insights from his work at ZML and lessons learned from putting inference into production at scale.

Inference Everywhere: optimizing performance

At Photoroom, building AI for creative tools means constantly experimenting and sometimes (pretty often actually) rebuilding from scratch. In this talk, Jon Almazan, Machine Learning Researcher at Photoroom, takes you behind the scenes of the team’s latest open-source text-to-image model, designed to push the limits of visual fidelity, controllability, and open reproducibility. He’ll share how the team iterated on training pipelines, data composition, and architectural tweaks to make the model both high-performing and publicly usable  and why open research remains essential for innovation in production environments. 

Expect a technical deep dive into what worked, what didn’t, and what it really means to build a new model when starting from zero.

Inside Photoroom’s Open-Source T2I model

Gain deep insights into their multi-agent AI architecture, observing how the system evaluates dynamic combat situations, formulates tactics, and executes maneuvers. The session will explicitly address the immense engineering complexities of bridging the gap from high-fidelity simulation to live flight operations. A technical leap akin to DeepMind's transition from virtual to physical agent control, but within the stringent demands of aviation safety and performance. Prepare for a technical deep-dive into real-life AI agent deployment at its most critical.

The first AI flying a production fighter jet

Exploring why language remains the highest expression of human intelligence, and how language translation is exposing the limits of today’s AI architectures and pushing us to build DVPS, a new open foundation model designed to go beyond them.

Translation and the limits of autoregressive transformers.

MCP has grown to be a key component for AI agents connecting AI with the world. 

We will dive into the latest development of the stack including MCP, ACP, and ChatGPT apps… Then explore best practices and tips on how to build, deploy and monetize an MCP server in production.

Central room

A deep dive into MCP and ChatGPT apps

AI is not a mere new vertical in consumer products. Rather, it is a technological shift that already applies to almost all layers of any product or business.

At Spotify, it means improving all of our core strategies:

- Ubiquity: AI-powered interfaces and our integration with inference providers.

- Personalization: Reasoned recommendations to get you to the next discovery, and a conversational agent as your personal DJ

-Freemium. Improving our monetization funnel that already drive hundreds of millions of paying subscribers, via novel experiences in playlist making or audiobooks discovery.

AI also presents new challenges as regards generated music and audio. We will share our approach to ensuring a trusted experience for both listeners and creators.

AI everywhere: reinventing Spotify's core strategies

When you think cooling, you often think fans and CRAC units. But for end-to-end AI stacks: from chip through server to data center, the thermal challenge spans multiple layers. 

In this session we’ll map the full spectrum: 

We’ll bring case studies, benchmark data, and methodology insight from companies owning each layer of the chain. 

Beyond Air Cooling: new frontiers in AI hardware design

Modern retail platforms rely on complex product taxonomies to power sourcing, merchandising, logistics, pricing, and search. But Veepee’s taxonomy : ∼1,400 nodes, evolving constantly, non-MECE and highly imbalanced, makes full automation challenging.

In this talk, Anis Gandoura (Principal AI, Veepee) presents how his team moved from a zero-shot CLIP baseline to a production-ready multimodal classifier that reaches or exceeds human-level accuracy.

- how zero-shot CLIP performs on real, high-noise retail data,

- how contextualizing category embeddings boosts performance,

- and how a multimodal LLM + “label book” achieved 94% Top-1 accuracy with no heavy training.

The result: a flexible, low-maintenance system that scales with taxonomy evolution, cuts manual corrections, reduces time-to-list, and improves downstream data quality.

Cracking Product taxonomy at scale: A multimodal, zero-shot approach

This panel contrasts two technological and conceptual models of creativity in the computational age.

Ircam Amplify representatives, Nathalie Birocheau and Frédéric Amadu, will outline systems which detect AI-generated artifacts in audio tracks through signal analysis, helping secure production pipelines and safeguard human creative processes.

Corina Chutaux, PhD in Digital Humanities will bring another perspective of creativity that includes computational agents, enabled by multimodal generative models and their capacity to produce novel aesthetic configurations.

Together, the speakers will examine what defines creativity in an "artistic work" today, and if and how human can remain at the center of creation in an age of generative sound.

Creative Integrity and Algorithmic Innovation: A Dual Approach

Distrust of AI systems remains the biggest barrier to adoption in high-stakes decisions. This session examines how agentic architectures address this through radical accountability, building systems where each AI agent handles one narrow, auditable task. Henrik Landgren, Co-founder and CPTO at Gilion and formerly Spotify's first VP of Analytics, shares how Gilion developed analytical tools that achieved 90%+ accuracy in 12-month revenue forecasting across €100M+ deployed, and how financial institutions can adopt these tools to supercharge their investment strategies. Just as GitHub transformed software by letting developers codify and version control their code, agentic AI enables investors to programme their "secret sauce" into analytical networks. While 2025 was the year for experimentation with agentic AI, in 2026 and beyond this technology is no longer optional for investors who want to stay competitive. Companies gathering investment data now build an insurmountable early-mover advantage and data moat.

Finally solving the high-stakes problems with agentic tree structures and metrics pipes

AI for video now spans two radically different architectures: 

Between them, broadcasters like France Télévisions must guarantee reliability, content provenance, and metadata integrity under production constraints.

This panel breaks down the engineering behind each layer. Inference pipelines, asynchronous rendering, edge vs cloud trade-offs, and how standards are reshaping video AI at scale.

A technical deep dive into how video AI systems are actually built, deployed, and trusted in production.

From edge inference to generative pipelines: building video AI that works

AI is shifting from intelligence to agency—as Andrej Karpathy puts it, intelligence is now abundant; agency is what’s scarce. Agency is the ability for systems to act, decide, coordinate tools, and pursue goals autonomously. This Wave of Agentification is transforming how companies operate: traditional work is declining, AI-assisted tasks have plateaued, and most organizations remain stuck wondering why 95% of AI projects fail.

This talk goes inside the agentic stack, sharing frameworks, metrics, and lessons from Mirakl’s journey toward Agentic Commerce. We’ll explore how to design, deploy, and scale agentic systems that unlock real autonomous value creation.

Inside the Agentic Stack: How Agency Is Rewriting Enterprise

Pharmaceutical R&D is being reshaped by AI. Not through buzzwords, but through tangible engineering breakthroughs.

Biolevate and Owkin have built AI pipelines that accelerate molecular discovery, predict biological activity, and optimize clinical research at scale. These systems blend multimodal data (genomic, molecular, imaging) with large-scale model architectures to guide scientists toward more effective treatments.

This session explores how these AI models are deployed in real research environments:

Moderated by Sophia Metz, this discussion will go under the hood of AI-powered drug discovery: the data, the models, and the systems making faster, smarter R&D possible.

From molecules to models: How AI is redesigning pharmaceutical R&D

For 30 years, video creation meant tools like Adobe or Canva: powerful, but manual. Then came AI prompting: you could just type what you wanted. But without structure, results often fell short.

Today, Aive takes us into a third era: video built from a brief, not just a prompt.

Their proprietary Multimodal Generative Technology (MGT) combines multiple AI models and meta-learning, framed by data.

After 6 years of R&D, MGT is here and it works! Join Alexandre Thomas, VP of Engineering at Aive, to open the hood of MGT and explore how this proprietary technology really works — a deep dive for experts eager to understand the mechanics behind next-gen video creation.

Founders Café

Decoding Aive's technology behind next-gen video transcreation: MGT

Inference speed is the bottleneck holding back the next generation of AI applications. Agentic workflows, real-time creation, and test-time compute scaling all demand much faster token generation at runtime.

This talk reveals Kog's path to 100x faster AI inference through systematic GPU optimization and novel Transformer architectural innovations that circumvent hardware constraints. When AI operates at 10,000 or 100,000 tokens per second, computers stop being static rule-execution engines and become smart, real-time adaptive platforms. Programming in natural language becomes practical. Complex games and applications generate themselves dynamically. Agents and Deep Research are instant.

Gaël, founder and CEO of Kog, will share the unique research insights and GPU engineering breakthroughs used to build the future of real-time AI computing

Real-time AI: building 100x faster inference

Designing a printed circuit board (PCB) is a classic example of a complex combinatorial problem: hundreds of components need to be connected without overlaps, while minimizing path lengths and respecting physical constraints. The number of possible configurations quickly becomes astronomical, making traditional approaches inefficient.

In this 15-minute demo, InstaDeep will showcase how Reinforcement Learning (RL), combined with their MCP framework, can tackle PCB routing by learning to “play the game” of placing and connecting components. Beyond PCB design, this approach illustrates how RL can be applied to real-world optimization problems in industries such as logistics, infrastructure, and even biology.

Solving complex combinatorial challenges with Reinforcement Learning: the PCB routing demo

Real-world perception for autonomous driving (AD) requires models that can learn from massive, uncurated, multi-sensor datasets with minimal, costly labeling. Self-supervised learning (SSL) is key to training foundation models in this regime. This talk moves beyond canonical SSL to present open-source foundation models from our team. We will highlight key examples, including Franca, our fully open-source foundation vision encoder (data, models, and code), and VaVIM/VaVAM, our open generative video and video-action models. We focus on how these models can be built on top of each other, creating a composable stack for perception. We'll show how this enables advanced applications like cross-sensor distillation, auto-labeling, and architecture repurposing, ultimately saving time, money, and improving performance.

This work exemplifies why open models are critical for developing the next generation of annotation-efficient and reliable autonomous systems.

Open & Repurposable: Foundation Models for the Automotive Industry

In this hands-on workshop, participants will explore how to transform their desktop environment into a powerful AI development environment using the GIGABYTE AI TOP ATOM—a compact "personal AI supercomputer.

From Desktop to Supercomputer: Building Local AI Workflows with GIGABYTE AI TOP ATOM

As demand for generative AI surges, so does data center power consumption. From edge to cloud, the challenge is clear: how do we scale AI sustainably?

This session presents a compelling answer. For AI inference — the majority of enterprise AI workloads — we'll demonstrate how to build infrastructure that delivers high performance, Total Cost of Ownership efficiency, and sustainability all at once. Enter FUJITSU-MONAKA: an Arm-based CPU leveraging supercomputer technology from Fugaku, the world's fastest supercomputer.

Through architectural deep-dives and benchmark comparisons against leading processors, we'll validate the viability of next-generation AI infrastructure. This is practical engineering for sustainable AI at scale.

Scaling AI Sustainably with FUJITSU-MONAKA

In this talk by Dr. Antoine Jardin, researcher at CNRS and CTO of Arlequin, we explore what it means for information to be true and authentic in the age of generative AI. Fake news and misinformation is an ever-increasing threat. Defining safety and authenticity requires deep understanding of how models are trained and developed, and even the smartest or safest of models may still lack the ability to differentiate truth from fiction. Using a combination of unsupervised learning and multi-modal inference, Arlequin helps cut through the noise and identify narratives without inferring bias from the analysis. We will explore how combining methodologies will lead to a safer, more auditable and understandable superintelligence.

Inductive learning, AI safety, and superintelligence in the age of disinformation

In this hands-on workshop, Adam Marszowski (Solutions Architect at SOFTIQ) will present a use case based on a real-world production deployment for an e-commerce client, showcasing how Computer-Using Agents can autonomously search the web, interact with diverse real-world interfaces, and generate qualified B2B leads without human intervention.

The session will demonstrate how to run such agents on Scaleway infrastructure, highlighting best practices for reliability, security, and performance at scale. Attendees will gain both a high-level understanding of how these systems operate and expert implementation tips drawn from a production-grade deployment.

Building computer-using agents for AI lead generation

While generative AI has revolutionized how we work with unstructured data like text, images, and audio, the most important business assets remain in structured formats like tables and spreadsheets. These datasets are central to decision-making, powering predictive tasks that optimize processes across the entire business chain, from sales and marketing to product and operations.

Yet, data science teams handling these tasks still rely on traditional machine learning (ML) methods, often spending weeks, or even months, developing, training, and maintaining models. This can significantly slow down collaboration with business teams and delay crucial decisions that depend on the models’ outputs.

In this talk, Alexandre Pasquiou, co-founder of Neuralk AI (Future 40, Station F) will present a new generation of frontier models designed to accelerate the data science life cycle: tabular foundation models (TFMs). He will reveal NICL 2.0, Neuralk’s powerful new tabular foundation model, capable of generating instant predictions for any ML task without the need for retraining, cutting weeks off the typical pipeline. He will break down the innovative principle of in-context learning, which allows NICL to adapt on the fly to any new dataset or task using examples (“context”) provided at inference time, similar to how LLMs respond to prompts.

Tabular Foundation Models: the new frontier of data science

In this hands-on workshop, participants will learn how to transform a simple photo of a handwritten restaurant menu into a structured, machine-readable inventory — the secret behind the flawless organisation of the event’s fictional coffee shop, Café de l’AI.

The session walks through each layer of a vision-driven workflow:

By the end of the workshop, attendees will have built a fully functional pipeline that turns an image into a restaurant inventory.

This structured menu will then serve as the knowledge base for the follow-up workshop, where participants build a voice-first AI agent capable of taking customer orders through WhatsApp.

Build the OCR-powered menu inventory for “Café de l’AI”

In this hands-on session, participants will learn how to build a fully functional audio agent capable of taking customer orders orally through WhatsApp, for the event’s fictional restaurant "Café de l’AI"

The workshop will walk you through each layer of the system:

- Audio transcription models to convert speech into structured, machine-readable text in real time.

- Structured output to interpret customer intent and translate them into actionable data.

- Model Context Protocol (MCP) to orchestrate communication between the agent’s reasoning engine and external tools or APIs.

- n8n for workflow automation, connecting the conversational layer with order management and backend systems.

By the end of the session, attendees will have built a voice-first AI interface combining speech recognition, natural language understanding, and automation.

Build the audio agent for “Café de l’AI”

Learn how to run AI on truly green infrastructure and why sustainability starts with accurately measuring impact.

Discover practical methods to track energy usage, present transparent impact data, and engage stakeholders effectively.

Explore tools, architectures and user education strategies that make AI systems measurably greener and more responsible.

How to build a sustainable AI infrastructure

Investing in Artificial Intelligence is now a key responsibility for any manager looking to innovate or add value to their products or services.

Comex frequently asks the same questions: What are the priorities? What is the return on this investment, and how do we accurately measure the ROI of AI?

What is the best strategy for launching an AI pilot and assessing its impact? While measuring, steering, and optimizing are essential actions, executing them within a framework of trust is even better.

How can you lead teams to generate rapid, high-impact results within a regulated environment focused on continuous optimization?

Agenda 2025