Dev Summit 2026

In a complex omnichannel environment like Decathlon’s, distinguishing between correlation and true incremental impact is the ultimate challenge. Traditional A/B testing often falls short when physical stores, local events, and global trends collide.

This session dives into our journey of scaling measurement through Randomized-Controlled Trials (RCTs) and Geo-experiments. We will explore how we utilize techniques to quantify the real value of business use cases, like testing the price elasticity of the products.

However, the real breakthrough lies in how we think about the evolution of this process. We will showcase our vision of Agentic AI: moving from manual statistical workflows to autonomous AI Agents that can help on the experiment design, select optimal control markets, and translate them into actionable business insights.

It's a great opportunity to see how we are democratizing rigorous data science to ensure every decision at Decathlon is backed by the "Truth of Causality."

AI is moving fast - and so are the decisions that come with it. For engineering teams working in complex, real-world systems, the challenge isn’t access to AI technology. It’s knowing where to start, what to build, and how to make the right choices across an increasingly sophisticated stack.
This talk shares practical approaches to identifying high-value AI use cases, drawn from experience designing and deploying production machine learning systems inside one of the world’s most demanding environments - global financial market surveillance across more than 50 exchanges and 15 regulators worldwide.
A core theme is the relationship between problem definition and technology selection. Before choosing a tool, understanding the process, the inputs and outputs, and what genuinely better looks like tends to produce stronger outcomes - and clearer architecture decisions.
The session also explores how classical ML, generative AI, and agentic systems can work together effectively. Rather than viewing these as competing approaches, the talk offers a framework for thinking about each as a distinct capability - and for combining them deliberately to build systems that are greater than the sum of their parts.
Attendees will leave with practical frameworks they can apply to their own environments, a clearer mental model for navigating the modern AI stack, and fresh perspectives on how some of the most complex AI systems in the world are being designed and deployed today.

Over two years of integrating AI into our client projects, we've dramatically accelerated some activities (and broken others). This talk shares what we've learned: where AI truly changes a designer's day-to-day, what works in production, and what we've stopped delegating to it.

## Part 1 : Stepping back: what has actually changed

How AI is transforming a designer's daily work:

- On their core activities
- On the organisation and infrastructure that supports them (ops and design systems)
- On their collaboration with other disciplines (developers and product managers)

We'll also look at how the evolution of models and solutions now makes it possible to bet on specific tools for specific use cases. We'll see that the learning curve has dropped significantly: recent interfaces and models make these tools accessible to non-technical profiles — and that's precisely what's shifting the dynamic between disciplines.

## Part 2 : Real use cases: AI in the workflow

Real examples from client projects such as Devialet, Bpifrance, Rexel and CMA CGM: prototyping, AI-augmented user research, functional specification generation and production. We'll show concretely how these tools fit into the day-to-day of a project: when we use them, on which types of deliverables, and what it actually changes in the way we work.

We'll also talk about the recent shift towards a central artefact — the code — where everything else (documentation, specifications) is now generated automatically by AI (flows in FigJam, functional specs in Notion, etc.). In this setup, we'll show how documentation artefacts can update the code directly, cascading changes through the entire chain.

## Part 3 : Lessons learned, risks and perspective

A synthesis of what these experiences have taught us — including the pitfalls. Notably:

- The "workslop" phenomenon: AI output that looks finished but is structurally broken
- Technical literacy making a comeback as a key designer competency
- What can't be delegated — and why we've walked back on certain use cases
- What we take away from all this: craft (judgement, structure, rigour) remains what separates usable output from throwaway output, and it's what repositions the designer as a competitive advantage

Innovative products rarely fail because of a single bad line of code or a flawed business hunch. They fail because of organizational silos. Product Managers, Developers, and Data Specialists often speak entirely different languages—and the glaring issues only surface on launch day.
This high-energy, 60-minute workshop flips the script. Using the Pre-Mortem framework, we won't ask you to build a successful product; we’ll ask you to dissect a complete disaster. Working in cross-functional, mixed tables (combining ADEO group and Decathlon people), participants will tackle a realistic, hypothetical launch. Business will have to face hard technical constraints, and Tech will have to answer to market realities.
The ultimate twist? The rescue plan your team builds must survive the final "Crash Test": a Generative AI acting as a ruthless "Devil's Advocate," programmed to expose the logical, data, or coding blind spots you missed.
Spend one hour learning how to fail safely in a room, so you don't fail spectacularly on the market.

Generative AI opens exciting opportunities to reinvent retail experiences. However, transforming a promising demo into a reliable product feature is far more complex than it first appears.

In this talk we will present how we moved from an AI demo built for a hackathon to a full fledged AI feature pushed to the millions of users of our mobile application.
Integrating image generation models into an existing application presents unique architectural challenges. How do you ensure the model quality despite the constraints of a mobile application? How do you balance costs with massive user scale?

We will trace the PaintMyWall journey from early AI experiments to introducing mobile capabilities, allowing customers to test colors on their walls with a simple photo, before taking a look at the opportunities this technology unlocks.

Last year, I navigated a unique challenge: working simultaneously across multiple teams with
polar-opposite leadership styles. PM-A was a short-term "speedster" focused on low-cost
experiments. However, PM-B is completely oppositie, investing half a year into a solution only to
find it brought no business value, with customers ultimately saying,
"This isn't what we wanted.

My analysis of our 2025 performance revealed a startling paradox: while 17 out of 24
successful A/B test wins(precise number needs to be checked) came from short-term fixes,
our product began to feel like a collection of disjointed patches. We often promise "we will
iterate it next quarter"

Now that we've handled the immediate issues, let's move on to other
priorities" or as usual other prio comes in the next quarter. We were winning the sprints, but
losing the vision.
This talk focuses on the two specific methodologies we used to reclaim our strategy: Design
Sprints for rapid alignment and ideation, and Amazon’s "Working Backwards" framework to
define the successful customer experience before writing a single line of code.

In this session, I share a 5-step framework to bridge the gap between
"now" and "later":

1. Vision Co-Creation: Combining Design Sprints and Working Backwards to build a
1-2 year roadmap that acts as a compass for every Jira ticket.

2. Goal Definition: Moving beyond conversion metrics to ensure every experiment has a
"Strategic Fit.

3. Continuous Discovery: Meeting customers regularly to separate short-term "noise"
from long-term "signals.

4 Experiment Alignment: Tactical ways to ensure small experiments are foundations for
the future, not just temporary band-aids.

5.The 70/20/10 Portfolio Rule: A resource model to protect team energy for Incremental
(70%), Evolutionary (20%), and Disruptive (10%) work.

Attendees will leave with a toolkit to transform from "pixel-pushers" into product strategists who
can balance immediate KPIs with a sustainable, long-term vision

I spent months architecting a complex SaaS for blockchain trading. I built robust microservices, complex Backend-For-Frontend (BFF) aggregation layers, and intricate state management systems just to feed a heavy web application. And then, a single autonomous agent (Openclaw) bypassed all that complexity, proving it could cover the exact same needs with extreme, user-driven personalization.

This experience was a wake-up call: **Autonomous agents are becoming the new major channel of the web.**

For companies building large-scale B2C platforms, this is a massive paradigm shift. Tomorrow's customers won't just browse our websites; they will send their personal AI agents to find products, interact with our services, and complete transactions for them. If our architectures are solely optimized for human screens, we will miss this revolution.

In this 45-minute deep dive tailored for developers, architects, and tech professionals building the future of our applications, we will explore how to evolve our systems to serve this new channel:

- **Beyond the BFF (Backend-For-Frontend)**: Why optimizing our APIs solely for screens is no longer enough. We will explore how to transition from screen-specific aggregators to exposing atomic, autonomous "Capabilities" (Tools).

- **Designing Agent-Ready APIs**: How to design self-describing capabilities optimized for LLM execution, memory, and reasoning, allowing them to coexist with your traditional frontends.

- **Protocols of the New Webv**: A technical deep dive into **WebMCP** (standardizing how systems expose tools dynamically without web scraping) and **UCP** (granting agents transactional capabilities).

- **Live Demo - "*From API to Skill*"**: What happens when a platform isn't "agent-ready" yet? I will demonstrate how to wrap a traditional REST API (using Strava data) and transform it live into a standardized, agent-ready "Skill" to instantly generate a highly personalized AI sports coach.

Join me to discover how to future-proof your architecture, bridge the gap between human UIs and AI agents, and start building powerful capabilities for the agentic web.

The Geo-Intelligence initiative represents a strategic shift in how market expansion opportunities are identified and prioritized. By leveraging the ArcGIS platform (a leading Geographic Information System - GIS, used to analyze, visualize, and manage spatial data), we developed an AI-driven, state-of-the-art geospatial framework that integrates spatial data, market indicators, and internal sales performance to support data-driven geographic decision-making.

The analytical architecture combines state-of-the-art statistical modeling techniques, geospatial data enrichment, and automated ZIP code geolocation, enabling accurate estimation of market potential and current market penetration at a granular region level.

This approach also supports marketing and commercial strategy, helping teams identify high-value territories, target campaigns more effectively, and allocate sales resources based on geographic opportunity.

In addition, the framework incorporates state-of-the-art predictive models for sales forecasting, allowing the organization to estimate potential revenue and prioritize expansion areas before investments are made.

Integrated within the ArcGIS ecosystem, this initiative transforms geographic analysis into a scalable decision-support capability, enabling faster, data-driven expansion strategies and improved revenue growth planning.

Chatbots have quickly become the default answer to conversational AI. They appear everywhere, sometimes useful, most of the time intrusive, and often represent the first idea teams reach for when exploring AI-driven interactions.

But conversational experiences extend far beyond chat interfaces and open text fields.

In this talk, we will explore how conversational design can be thoughtfully integrated across products and services. Drawing from real implementations at ADEO, we will share the design principles guiding our work, the case studies that shaped these guidelines, and the cross-disciplinary processes that help designers, product managers, and engineers deliver meaningful, coherent AI-powered experiences.

**Every time a search engine displays a product, it makes a decision**. Not just a sort order, but a business decision that impacts customer experience, revenue, and fairness between competing products and channels. But what actually defines a good decision?
**A ranking system is effectively a judge: every query is a trial, every product is a candidate, and user behavior becomes the evidence**. The challenge is that this evidence is incomplete, biased, and distributed across channels. Clicks measure attention but not intent, while conversions capture purchases but miss delayed or offline behavior.
Search ranking must also balance competing objectives. The products people click are not always the ones they buy. Visually appealing items attract curiosity and traffic, while the products that generate revenue may appear less attractive at first glance. Optimizing purely for purchases can also favor high volume, low cost items and erode overall business value.
Time adds another layer of complexity. Some products perform consistently year after year, while others spike during seasonal moments or emerging trends. Ranking systems must decide whether to trust long term reliability or react to short term signals.
At the same time, behavioral signals are distorted by position bias, sparse data, and delayed attribution across channels. **In an omnichannel environment where customers research online and purchase later in store, the ROPO (Research Online Purchase Offline) effect makes naive metrics misleading**.
To address this challenge, the Search and Publication team developed a decision pipeline that industrializes the production of implicit judgments and acts as a **Truth Provider for AI systems**. The system transforms noisy behavioral signals into stable implicit judgments used to evaluate ranking systems, guide experimentation, train machine learning models, and provide reliable ground truth signals used by multiple internal products.
Built on top of BigQuery and dbt, the pipeline processes hundreds of millions of search sessions and several millions of products every day. It integrates multi channel evidence, bias correction, Bayesian priors, uncertainty estimation, temporal weighting, and composite indicators balancing discovery and purchase signals.
Attendees will gain a practical understanding of how to transform noisy behavioral data into reliable implicit judgments, and how to choose the right signals when discovery, conversion, and long term reliability pull ranking decisions in different directions.
**Ranking is not just an algorithm problem. It is a decision problem**. Every decision needs an honest definition of what good actually means.

The idea is to show an MCP server running with our performance testing tool at ADEO global scale: OctoPerf.

The aim is to further democratize the implementation of performance testing, but above all to save considerable time in analyzing reports and improving the quality of script writing.

The idea is to show that, with AI, each team will now have an agent capable of communicating with our server and acting like a performance test engineer.

In the world of global retail, balancing product availability with inventory cost is a high-stakes game of precision. For years, managing "Safety Stock"—the buffer inventory held to mitigate demand and supply uncertainties—at Decathlon relied heavily on empirical rules, based on fixed numbers of coverage days regardless of the product, inducing high-level coverage values. While functional, this "one-size-fits-most" approach lacked the granularity needed to truly optimize a complex network of thousands of stores and unique items.

This session details the technical evolution of our Safety Stock solution, moving from these empirical rules to an automated, industrial-scale statistical engine. We will present a deep dive into how we built a system that computes dynamic safety stocks at the Item x Store x Day level.

Key technical pillars of the session include:
- The Statistical Engine: Moving beyond the baseline to formulas integrating forecast error variability, and lead time variability to meet specific target service levels.
- The Architecture: An overview of our data pipeline, utilizing dbt for weekly computations based on Decathlon's datalake and Talend flows to push granular safety stock quantities directly into SAP F&R.
- Backtesting & Impact: A transparent review of our pilot results across pilot product families, demonstrating a reduction in Days Sales of Inventory (DSI) of a few days without sacrificing availability.
- Product Segmentation: How we use ABC-XYZ classification to automatically tailor service level targets based on both business targets and demand predictability.

Join us to learn how we bridged the gap between supply chain theory and real-world data science and analytics engineering, transforming inventory management from a "best guess" into a precise, automated competitive advantage.

Securing our digital products is a major challenge. Cybersecurity analysts often lose precious time dealing with time-consuming processes and sometimes incomplete tools. This slows down the identification of risks and vulnerabilities, as well as the remediations that development teams need to apply.

To tackle this issue at ADEO, we decided to rethink our approach: what if we designed a squad of specialized AI agents ?

In this talk, we invite you behind the scenes of this initiative. You will discover our goals, our initial results, and how this multi-agent architecture could ultimately transform your team's day-to-day work.

In an increasingly mobile-first world, delivering a seamless, high-performance shopping experience is essential.
Join experts from Leroy Merlin / ADEO and Decathlon for an exclusive panel discussion on the evolution of their mobile applications.
During this session, we will dive into the unique challenges of managing large-scale e-commerce apps, from enhancing user engagement and optimizing conversion funnels to integrating omnichannel features that bridge the gap between digital and physical stores.

Today, companies are producing more dashboards than ever before. Yet, many organizations find themselves with what I call "dashboard debt": dozens of screens, rarely used, and seldom linked to actual decisions.

We often hear that companies want to be data-driven. But in reality, many are becoming dashboard-driven instead. In this talk, I will defend a simple idea: stop delivering charts, and start delivering decisions.

The problem rarely lies with the charts themselves. It comes from what happens beforehand: a lack of discovery, requirements that are accepted too quickly, and solutions that are decided upon before the problem is even understood. When everything becomes a KPI, nothing is truly "key" anymore. And when an organization has 3 to 10 dashboards per employee, it's generally not a very positive sign of data debt management.

I will show how data teams can shift from a logic of producing dashboards to a logic of creating data products designed around real needs, thanks to a structured process of discovery, prioritization, and design. The goal is clear: fewer dashboards, but much more value.

Discover how to use a multi-agents strategy to improve and accelerate Product Management processes, from Strategy to Delivery.

We will share a framework and tools ready to use, as well as REX and usage examples.

The era of monolithic LLM assistants is giving way to a more powerful, resilient reality: Agentic Ecosystems. In this session, we will explore how to transition from isolated AI tools to a modular, 'Lego-like' architecture of autonomous agents that reason, plan, and collaborate.

We will dive deep into the technical 'glue' that enables this synergy, focusing on ADK for orchestration, and the critical role of protocols like MCP (Model Context Protocol) and A2A (Agent-to-Agent) for seamless communication. You will learn how we differentiate between specialized sub-agents—focused on niche expertise like data or validation—and the 'Coordinator' agent that acts as the system’s brain to plan and validate results.

Using INES (Our Digital Instore Assistant) as our primary case study, we will demonstrate how this ecosystem provides high-value operational efficiency by simplifying access to information for store teams. We’ll share our lessons on how to work on an independent agentic ecosystem.

Generative AI is no longer just a "Lab" project, it’s in production. But when you move from a playground to an industrial environment like Decathlon, the stakes change. Suddenly, "Prompt Injection" isn't a fun riddle; it's a potential data breach or a compliance nightmare.

In this session, we (PIC Data Team & Cybersecurity Team) share our raw, unfiltered journey of securing two major GenAI applications: a global translation engine and a content compliance engine.

We didn't just read the OWASP Top 10 for LLMs—we lived it through three rigorous professional pentests. We will walk you through the actual vulnerabilities the ethical hackers found and, more importantly, the concrete defense layers we built in our GenAI services to neutralize them.


Speakers :
Pierre-Yvan Devos (Data & AI Manager)
Karl Mongosso (Senior Machine Learning Engineer)
Philippe Lasek (Information Security Officer)

Opentelemetry has recently skyrocket to become a new standard in telemetry collection. Both Decathlon and Adeo have started incorporating it into their observability stack with different approaches. In this talk we will cover the following points :

- What is Opentelemetry ?
- A comparison between the approaches used by adeo and decathlon
- The opportunities of opentelemetry (perspectives from decathlon and adeo)
- Key results from both companies

At Decathlon, our Pricing Decision Makers were suffering from *infobesity*. We solved that by changing their UI from 12 complex Tableau dashboards into a **frictionless, natural-language AI assistant** integrated directly into Google Chat.

But taking a **Generative AI agent** from a **shiny POC** to a highly accurate robust **production** product is paved with technical hurdles.
In this talk, we will **open the hood** and share our engineering learnings.

LLMOps, monitoring, evaluation, costs ... Join us to discover what it actually takes to build and **run GenAI at scale**!

In an era where product complexity is rising and customer expectations are higher than ever, traditional training often falls short of the "real-world" pressure of the sales floor. Explore how AI Virtual coaching is bridging the gap between technical product knowledge and expert consultation.

This talk introduces an innovative Customer AI-Simulator for sales advisors co-built with sellers experts from LMIT and learning teams, and approved by UX. By leveraging Speech-to-Text and AI Dialogue, we have created a safe, interactive environment where sales advisors can practice complex scenarios, such as the technical "Shower Enclosure" family, with virtual customers to master their knowledge and be able to traslate produt characteristics in benefits for the customer.

Longer description:
The Problem that solution solves
- New advisors struggle to internalize the Leroy Merlin sales model.
- Senior advisors know the specs, but often struggle to translate them into customer benefits.
- Classroom training is disconnected from the daily reality of the store floor.

The Solution: The AI Virtual Coach MVP
- We’ve built a Customer AI-Simulator. This isn't just a chatbot; it’s a high-fidelity 'flight simulator' for retail. - a verbatim from a seller that co-designed this coach.
- Real-time Interaction: Using Speech-to-Text and AI Dialogue, advisors engage in a natural verbal flow with a virtual customer.
- Realistic Scenarios: From "budget-conscious" to "aesthetic-oriented," the AI mimics real human behaviors and objections.
- Immediate Feedback Loop: After the 'sale,' the AI provides an instant debrief based on LMIT standards—identifying strengths, missed opportunities, and discovery depth.

Why it’s a Game Changer (The Value)
- Risk-Free Practice: Advisors can fail, learn, and iterate in a safe environment.
- Standardization: Aligns every store to the same high-quality LM selling model (M2, M3, M4).
- Data-Driven Growth: We finally have visibility into how our teams sell, not just what they know.

Data models are often designed around data availability and development convenience, but not always around how users actually consume data. As analytical products scale, this misalignment can significantly increase infrastructure costs.
In this talk, I will share the transformation of the Seller Deep Dive Dashboard, which at one point became one of the most expensive dashboards to operate due to its data model design.
The initial architecture relied on a fully denormalized table, where dimensions were merged directly into the fact table powering the dashboard. While this simplified early development, the model became increasingly inefficient as data volume grew and slowly changing dimensions expanded the dataset.
By analyzing which KPIs users actually queried and how frequently they were used, we redesigned the model around a usage-driven star schema.
The new architecture separates metrics into two fact tables:


A core fact table containing roughly 80% of the most frequently used KPIs, representing only 20% of the total data weight


A secondary fact table containing the remaining less frequently used but heavier metrics, queried only when required


This change dramatically reduced the amount of data scanned by the majority of queries.
As a result, the dashboard evolved from being one of the most expensive dashboards to run, costing roughly €3500 over two weeks, to becoming one of the most efficient, reducing costs to approximately €350 over the same period—a 90% reduction in FinOps costs while maintaining full analytical capability.