Dev Summit 2026

From routine incidents to national crises, the France Cybersecurity team train our companies for the worst.
Join us as we break down our training methodology, share key results, and reveal the most frequent attack patterns we face today.
We’ll dive deep into a real-world scenario at Leroy Merlin France: after 2 hours of total network isolation, discover how store teams successfully maintained self-service and kept the store open.
///
La direction cybersécurité France entraîne nos entreprises à faire face à des situations cybersécurité allant de la réaction à un incident basique à une crise nationale.
Dans ce Talk, nous abordons comment nous mettons en place ces entraînements quels sont les résultats et les scénarios d'attaques qui sont le plus observés et sur lesquels nous devons être prêt.
Nous ferons un retour plus en détail sur l'expérience vécu en Magasin Leroy Merlin France après une isolation totale du réseau pendant 2 heures -> comment les équipes en magasins ont pu maintenir le libre service et ouvrir le magasin.

Deploying custom MCP servers requires more than just a connection—it requires a robust, governed perimeter. This session explores how we leverage our AI Gateway / Developer Portal as the central nervous system for all MCP traffic, transforming standalone LLMs into actionable enterprise tools.

We will do a deep dive into:

* The "Gold Path" Deployment: Utilizing the Developer Portal to publish and manage consumers.
* REST API Synthesis: Strategies for wrapping our existing internal REST ecosystem into MCP tools, allowing LLMs to interact with legacy and microservice architectures securely.
* The AI Gateway Architecture: A breakdown of the security primitives—including identity propagation, request filtering, and automated rate limiting—that protect our data.
* Deep Observability: Utilizing our telemetry stack to ensure every AI-driven action is logged, audited, and debuggable in real-time.

AI looks clean in tutorials. In production, it rarely is. This workshop offers a practitioner’s walkthrough of real-world ML development - exploring what the process actually looks like when you’re building detection systems for real clients, with real data, under real constraints.
Drawing on experience developing and deploying machine learning systems inside global financial market surveillance, the session is structured around three transferable lessons: how to translate a client’s business problem into a well-formed ML problem; how to stay flexible and recalibrate when early assumptions don’t hold; and how to navigate the statistical realities of imbalanced, dependent, and noisy real-world data.
Each lesson is grounded in real project experience - honest about the complexity, and focused on the thinking and decision-making that determines whether a model earns its place in production. Attendees will leave with practical insight they can apply to their own projects, and a clearer picture of what separates ML that ships from ML that stalls.

The "Post-Cookie" era has left the digital ecosystem at a high-stakes crossroads: sacrifice measurement and user experience, or sacrifice privacy. While current workarounds like "Clean Rooms" and "Contextual Targeting" offer temporary relief, they fail to bridge the gap between **total user anonymity** and **hyper-relevant engagement**.

This talk introduces **Fully Homomorphic Encryption (FHE)** as the ultimate architectural pivot—a cryptographic framework that allows for both precise attribution and real-time website personalization without ever decrypting user data. We will explore how FHE enables "Blind Personalization," where a server can rank products and serve tailored content to a user it doesn't "know," and how it solves the attribution loop with mathematical certainty.

By moving past the "FHE is too slow" myth, we will demonstrate how 2026’s hardware acceleration and edge computing are making **Zero-Knowledge Advertising**—from the first click to the final purchase—a production-ready reality.

From Jeremie Forgeon, Director of Product @Decathlon Digital

LMDIS (Last Mile Delivery Intelligence System) is an AI-powered decision support tool for daily delivery fleet planning at Golilla. The system automatically analyses ~300 delivery addresses overnight, combining Google Street View imagery, OpenStreetMap road topology, and Claude AI vision to generate a structured operational briefing per order.
At its core, LMDIS resolves two simultaneous constraints: the minimum vehicle size imposed by cargo characteristics (weight, dimensions, product type — sourced via TMS API) and the maximum vehicle size permitted by road accessibility (carriageway width, turning radius, physical restrictions). The intersection determines the optimal vehicle assignment; when no intersection exists, the system identifies the nearest accessible point for the largest available vehicle and flags the conflict for manual resolution.
The road analysis engine follows actual OSM way geometry to sample points along both the delivery street and the access route from the nearest primary artery — eliminating the linear coordinate offset errors that make naive Street View analysis unreliable. A four-level heading cascade (OSM geometry → Google Roads snap-to-road → panorama inference → manual review flag) ensures no analysis proceeds with an incorrectly oriented image.

Most retail models fail by confusing correlation with causality. High-value customers might create more projects, but does creating a project actually increase their value? This presentation showcases the Adeo XP Award-winning "Drivers of LTV" project, which isolated the true incremental impact of driver—from life events to digital interactions—to build a high-precision business simulator.

In an era of infinite output, where AI can generate endless solutions in seconds, the role of the human leader shifts from "creator" to "framer." I want to explore why the abundance of AI-driven production makes the quality of the initial input—the problem framing—the only true competitive advantage and the skill of the future combined with strategic alignment among product teams.
Product teams success will heavily rely on radical team alignment with shared outcomes. When the speed of outputs accelerates, the direction and strategy we are heading towards needs to stay clear and precise.
Let's move beyond "more" and master "better" with intentional input and unified outcomes and visions.

This keynote will explore how Zalando positions itself as a tech company, with AI at the core of its strategy.
It will highlight how AI is leveraged to elevate customer experience, driving differentiation through personalization, inspiration, and seamless journeys.

## Context

Leroy Merlin is more than a home improvement retailer; it is a **one-stop shop that empowers people with end-to-end solutions to enhance their homes and their lives**. Leroy Merlin Brazil is a benchmark within the group when it comes to service sales.

However, throughout our digital journey, services had historically been treated as a **complement to product purchases**, confined to checkout flows or hidden behind low-visibility pages.

This structural constraint reduced offer visibility, limited the sale of standalone services, and prevented us from fully leveraging one of our **core strategic assets**.

In this presentation, we will share how we **redesigned and rebuilt the digital service journey** to transform installation services into **searchable, discoverable, and scalable digital products**.

---

## Redesigning the Digital Service Journey

The initiative was grounded in a **robust Discovery phase**, including:

- Deep dives with the **Instala operational team**
- Interviews with **store floor staff at the Tietê store (São Paulo, Brazil)**
- **Behavioral analysis using Microsoft Clarity**
- **A/B testing experiments**

Based on these insights, we redesigned the user journey to make services **visible and accessible across the entire digital funnel**.

### Key Experience Changes

**Smart Search**
Relevant installation services are now suggested **directly within the search bar**, allowing users to discover services earlier in their journey.

**Listing Pages (PLP)**
Dedicated tabs allow users to **toggle between "Products" and "Services"**, enabling easier discovery of standalone services.

**Product Detail Page (PDP)**
A new **service component** was introduced on product pages, including:

- Transparent service pricing
- Warranty information
- A **toggle switch** enabling users to add installation services **directly to the cart together with the product**

---

## Preliminary Results

Early results already demonstrate the impact of integrating services directly into the product discovery journey.

The new experience generated a **242% increase in sessions that add services to the cart**, growing from an average of **40 to 137 daily visits** with service attachment.

We estimate this increase represents an **additional annual revenue of approximately R$613,000 from services attached to product purchases**.

Service discoverability also improved significantly. In just **five days after launch**, visits to **service pages increased by 17%**, demonstrating the impact of making services more visible within the search experience.

These early results reinforce our core hypothesis: **when services are made visible and integrated into the decision journey, they unlock new value within digital retail ecosystems.**

We call ghost stockout the situation in which the stock is marked as available on digital stock management inventories, but not physically available in the store.

We present an unsupervised algorithm for detecting ghost stockouts using only daily sales data. We model each SKU's daily sales outcome as a time dependent Bernoulli process with a parameter representing the probability of selling one product on a given day, denoted as p₀(t). A ghost stockout manifests as an anomalous zero sales streak, whose probability

P(N,t) = ∏ from i=1 to N of (1 - p₀(tᵢ))

serves as a detection signal. In a ghost free environment this signal is uniformly distributed across SKUs; in the presence of ghost stockouts, the empirical distribution exhibits a relative abundance of low P(N,t) values far exceeding the theoretical uniform baseline, a signature that is both theoretically predicted and experimentally confirmed.

We estimate p₀(t) via an adaptive moving average procedure that self calibrates its window size to maintain a bounded relative estimation error across the full range of SKU sale rates and seasonal profiles. Crucially, the entire pipeline, including p₀(t) estimation, error bounding and signal computation, was implemented in pure SQL on BigQuery, eliminating the need for dedicated ML infrastructure and dramatically reducing deployment cost.

Validated across two retail stores over one month (3,220 physical inventory checks), the method achieved an overall true positive rate of 15.5%, rising to 17.8% on self service departments. In the top performing department store pairs the TPR reached 35%, demonstrating that when applied to structurally well defined availability contexts, this fully unsupervised algorithm delivers detection performance competitive with supervised approaches.

You collect your users' feedback in Condens during your research, but feel like the tagging is an endless task.
You've been assigned to a research started by someone else and struggle to explore data you didn't collect.
You simply look for ways to pace up your UX research workflow.
This demo will explore ways to fast-track your analysis using assisted tagging (within and between sessions) and automate the clustering of feedback to improve your research delivery.
Come visit us if you want to know more

In the retail industry, the accuracy of product to concept mapping is the backbone of customer experience. At Leroy Merlin France, this represents a massive operational challenge: over 5.7M product-concept pairs (out of 67M) currently show poor semantic alignment, leading to misclassified products. Furthermore, at search time, 15.8% of the customer terms remain unmapped to our Search & Pub Knowledge Graph, inducing lower recall and precision. Solving these issues are a massive business opportunity.
While dense embeddings became the standard about semantic alignment, they often lack the lexical precision needed to fix these specific business gaps. This led us to explore an alternative: **Sparse Encoders**, a solution designed to bridge our massive product catalog with the explicitness of Knowledge Graphs to reclaim these lost data points.
We will walk through the evolution of our approach: from traditional term expansion (SPLADE) to a custom model with a key innovation: **using KG concepts themselves as tokens**. I will share the "behind-the-scenes" of our experimentation, including:
* **The Technical Pivot**: Why we moved from generative Seq2Seq models to Masked Language Modeling (MLM) and finally settled on a Distillation strategy.
* **Implementation Hardships**: Deep dive into the challenges of GPU optimization, custom DataCollators, and refining loss functions to enforce sparsity without losing semantic depth.
* **Real-world Results**: How we achieved 99%+ sparsity while maintaining alignment for complex product categories.

Beyond the code, this talk is a feedback loop on **data quality and model refinement**. Whether you are working on search, recommendation or data classification, you will gain practical insights into leveraging Sparse Encoders to build more robust and interpretable data products.

This initiative introduces a robust, Python-based Generative AI solution engineered to safeguard a 18-million SKU marketplace from regulatory and brand safety risks. By automating the detection of non-compliant products and offensive content, this tool replaces inefficient manual keyword searches with a highly scalable, semantic validation engine.

The Challenge: The "Scale vs. Safety" Gap Following a massive catalog expansion in 2025, the organization experienced a surge in "Out of Scope" (OOS) listings. While illicit goods are centrally managed, two critical operational gaps emerged:

- Strategic Misalignment: Non-core products (e.g., smartphones, toys, or food) infiltrating a specialized DIY/Home Improvement marketplace.
- Brand Safety: Offensive content and hate speech hidden within product titles and descriptions.
- Inefficient Current State: Local Business Units (BUs) relied entirely on manual, unscalable keyword searches within the Mirakl platform to police the catalog.

The Solution: Hybrid AI Auditing Architecture To address these challenges, we developed a dual-approach architecture combining automated catalog screening with ad-hoc auditing tools:

1. Proactive Pipeline (Catalog-Wide Screening): A comprehensive, custom-built tool designed to validate the health and compliance of the entire country-level catalog, encompassing both published and pending (pre-publication) selection. This background process operates through a robust four-step automated workflow:
1.1 Initial Selection Filtering: Executes a broad sweep across the entire BU and supplier catalog, utilizing Regular Expressions (RegEx) and country-specific rules to rapidly identify and isolate products that are potentially Out of Scope (OOS).
1.2 Semantic Context Analysis: Leverages Vertex AI to conduct deep Natural Language Processing (NLP) on the isolated items. It evaluates full product titles and descriptions to understand their true semantic context (e.g., accurately distinguishing a prohibited "smartphone" from a permitted "smartphone repair tool").
1.3 OOS Classification: Categorizes the analyzed products with high precision, definitively flagging non-compliant items and separating them from legitimate, brand-safe merchandise.
1.4 Automated API File Generation: Seamlessly operationalizes the decisions by automatically generating ready-to-upload files (via API integration), eliminating manual data manipulation and enabling immediate catalog updates.

2. Reactive Pipeline (Targeted Auditing App): A custom Streamlit-based web application that empowers non-technical users to conduct deep, ad-hoc catalog audits through a four-step automated workflow:
2.1 AI Query Expansion: Leverages Gemini (Vertex AI) to dynamically expand a single seed keyword (e.g., "Barbie") into a comprehensive cluster of related terms (e.g., "dolls", "toys") to capture evasive listings.
2.2 Live Search Analysis: Audits front-end webpage results, capturing and evaluating exactly what the customer sees in real-time.
2.3 Semantic Validation Engine: A hybrid verification step combining BigQuery product data with Gemini’s Natural Language Processing (NLP) to ensure the product aligns with the BU’s strategic scope.
2.4 Operationalized Outputs: Automatically generates actionable files, including a Mirakl-formatted auto-rejection list for zero-GMV items, alongside a prioritized manual review queue for high-risk, revenue-generating products.

Impact & Current Results: The implementation has successfully transitioned the business from reactive compliance to proactive risk mitigation:
- Scale at Speed: Processes 274 SKUs per minute (approx. 3,300 SKUs in 12 minutes).
- Proven Output: To date, the system has screened 178,000+ products, analyzed 1,641 risk keywords, and successfully delisted 46,000+ non-compliant items.
- Precision & Revenue Protection: Maintains a False Positive Rate (FPR) of

How do you build a conversational AI platform that combines capabilities from multiple teams across an organization?

In this talk, we share how we designed, scaled, and governed a multi-assistant AI platform, from first integration to a reusable governance framework. Through concrete REX with Opus and Locus teams, we'll show how we went from case-by-case integrations to a clear model where any team can plug their AI capabilities into the platform.

We'll cover architecture decisions, integration patterns, and the prerequisites we now require before anything reaches production.

Designing a data architecture that can handle the **massive volume of a retail operation** requires more than just good tools; it requires **hardcore engineering rigor**.

At **Obramax**, we completely redesigned our data infrastructure, moving away from **isolated legacy databases** to a **unified, event-driven model** hosted on the **ADEO Group's DFDP platform**.

This session dives deep into the **engine room of our new Data Analytics platform**.

We will demonstrate our **aggressive automation approach** to building a **Medallion Architecture**.

You will see how we use **Apache Kafka** to capture **high-volume events** *(from SAP, VTEX, etc.)* and land them in our **Bronze layer**.

From there, we will showcase our strict **GitOps workflow using Terraform and dbt (Data Build Tool)** to automatically refine data into the **Silver layer**, and finally aggregate it into **business-ready Gold layer datasets**.

> Come see how we **automated the entire lifecycle of a Kafka event**
> from **system integration** to **BigQuery dashboard**
> with **zero-touch provisioning and maximum scalability**.

---

# 🎯 Key Takeaways

Participants will learn:

### ⚙️ Event Streaming Data Platform

A deep dive into building a **Medallion Architecture (Bronze, Silver, Gold)** using **event streams on Google BigQuery**.

---

### 🧱 Engineering Data Pipelines at Scale

How to combine:

- **Apache Kafka**
- **dbt (Data Build Tool)**
- **Terraform**

to build **bulletproof, scalable data pipelines** on the **ADEO DFDP platform**.

---

### 🔄 GitOps Data Infrastructure

Practical **GitOps strategies** for achieving:

- **zero-touch data infrastructure provisioning**
- reproducible data pipelines
- automated platform governance

---

# 🏗 Architecture Themes Covered

- Event-Driven Architecture
- Apache Kafka
- dbt (Data Build Tool)
- Terraform
- GitOps for Data Platforms
- Google BigQuery
- Data Lakehouse Pipelines

---

# 👨‍💻 Target Audience

This talk is ideal for:

- Data Engineers
- Platform Engineers
- Integration Engineers
- Data Infrastructure Teams
- DevOps / Platform Engineering Teams

---

# 🧭 Why This Talk Matters

Modern data platforms require **engineering discipline at scale**.

Retail environments generate **massive volumes of operational events**, and traditional batch pipelines cannot keep up with the **speed and reliability required for real-time analytics**.

By combining **event streaming, infrastructure-as-code, and GitOps practices**, Obramax built a **fully automated data platform** capable of transforming raw operational events into **business-ready analytical datasets**.

This session shares the **engineering patterns and automation strategies** that make this architecture scalable and maintainable.

---

# 🏢 Case Study

**Company:** Obramax

**Architecture Stack:**

- Apache Kafka
- ADEO DFDP Platform
- Google BigQuery
- dbt (Data Build Tool)
- Terraform
- GitOps Workflows
- Event-Driven Data Pipelines