talk-data.com

Seizième édition, même ADN : du concret, zéro bullshit.

Cette fois, on se penche sur la façon dont la génération IA et les approches agentic redéfinissent le produit.

2 talks + 1 lightening talk.

1️⃣🎙 Scaling Data Self-Service : 3 ans d’apprentissages chez Back Market Speaker : Capucine Delecourt – Group Data Product Manager @ Back Market

Ce retour d’expérience retrace les différentes étapes de la mise en place du self-service data chez Back Market : de l’approche initiale basée sur les formations, à une vision plus produit centrée sur les besoins utilisateurs. On y évoque les succès comme les limites, les difficultés rencontrées autour de l’outillage (LLM, outils de data viz), l’importance des fondations techniques (documentation, gouvernance, usage réel), ainsi que les ajustements nécessaires face à la montée des coûts. L’objectif est de partager des enseignements concrets sur ce qu’implique réellement une démarche de data self-service à l’échelle.

👩 Speaker: Capucine Delecourt est Group Data Product Manager chez Back Market. Après 7 ans de conseil en data product management auprès de nombreuses entreprises, où elle a pu explorer des sujets variés allant du machine learning à l’analyse produit, elle rejoint Back Market il y a trois ans pour piloter une équipe de PM dédiée aux enjeux data. Avec son équipe, elle travaille notamment à la construction d’une plateforme data scalable, au développement du self-service, et à l’intégration des bonnes pratiques produit dans un écosystème résolument technique.

2️⃣🎙 Les agents IA peuvent-ils remplacer une équipe Scrum ? Speaker : Renaud Chevalier – CTO @ Thiga

Afin de répondre à cette question, nous avons mené une expérimentation en 3 étapes testant les capacités d'IA « facilement accessibles » aujourd'hui : 1. Un assistant personnel avec un Custom GPT OpenAI 2. L'automatisation de workflows avec n8n 3. Des agents autonomes avec Flowise

Conclusion : L'IA ne remplacera pas les équipes Scrum, mais elle balaye tous nos repères (collaboration, rôles, artefacts, événements, pratiques, méthodes…). Nous entrons dans une ère de création produit radicalement différente, où paradoxalement l'agilité devient plus cruciale que jamais.

Une session accessible qui ouvre le débat sur l'avenir du travail d'équipe à l'ère des agents IA.

🎤 Speaker: Renaud Chevalier En tant que CTO et Agent Builder chez Thiga, je conçois des architectures d'Agents IA qui transforment la façon dont les équipes Produit travaillent et créent de la valeur. De la R&D aux implémentations en production, je teste et industrialise les approches qui amplifient réellement l'impact business.

⚡️🎙 Reimagining UX with Agents Speaker : Naji Alazhar – Cofounder @ Umans AI

En développant la plateforme multi-agents umans.ai, trois techniques nous ont bousculé les repères UX. Nous expliquerons ce que c’est, comment on le fait, et les opportunités créées :

• Remote agents + UI générative : La conversation génère l’interface à la volée, pendant que des agents distants exécutent des actions : l’utilisateur reste dans son flow, les clics disparaissent.
• Recherche Agentic : Dans une recherche augmentée, l’agent relance la recherche, ajuste la question au contexte et ramène une réponse plus ciblée.
• Mémoire d’agent : Un agent qui se souvient de vous peut nettement améliorer la qualité des interactions.

Venez challenger ces pistes et nourrir la réflexion avec vos propres expériences.

Plus d'infos à suivre...

We are pleased to invite you to our upcoming meetup co-organized with Artefact on April 16 at the Artefact Utrecht Office. Get ready for a deep dive into the AI-assisted programming for data engineers: beyond the vibe coding hype. Wanga Mulaudzi will share her journey from astrophysics to AI-powered code migration, while Adithya Krishnan will unravel the mysteries of context in AI-assisted tooling. Whether you're into space, code, or just curious how AI thinks, this one’s for you. 🚀💡

SCHEDULE 18:00-19:00: Welcome with food and drinks! 🙌 19:00-19:45: Talk 1 - Artefact - GenAI for Code Migration: My Journey From Supermassive Black Holes to Supermassive Codebases 19:45-20:00: Break 20:00-20:45: Talk 2 - MotherDuck - Hold on, where's my context...? 20:45-22:00: Networking / drinks!

TALKS [Talk 1]: “GenAI for Code Migration: My Journey From Supermassive Black Holes to Supermassive Codebases" by Wanga Mulaudzi

This talk explores the parallels between problem-solving in Astrophysics and tech, with a focus on Generative AI. From simulating the first image of the M87 black hole as part of the Event Horizon Telescope collaboration to developing AI-powered tools for code migration, we’ll dive into the challenges of computational modeling, automation, and large-scale migration of legacy code. We'll discuss the transition from academia to tech and examine how GenAI accelerates automation—contrasting the manual, high-performance computing workflows used in Astrophysics with GenAI now revolutionizing code migration. Expect insights into black holes, Python-driven automation, and lessons learned from tackling complex problems at the intersection of science and technology.

[Talk 2]: “Hold on\, where's my context...?" by Adithya Krishnan

Much of AI-assisted tooling depends significantly on obtaining the appropriate context for the specific task. But how exactly do these AI tools retrieve and utilize this context? And how can you, as a user, effectively provide and work with this context? That's exactly what I'll be talking about :)

DIRECTIONS

Address: Creative Valley building. Stationsplein 32, 3511 ED, Utrecht.

Directly next to the Utrecht central train station, you’ll find the Artefact office in the Creative Valley building.

Mark your calendar for the next session of the PyData Paris Meetup, on April 8th 2025. This Meetup will be hosted by Artefact, a global leader in data and AI consulting services, in their offices, 19 rue Richer 75009 Paris. The speakers for this session will be Martin Renou, Abdoulaye Sakho and Vincent Auriau.

Schedule 6:45pm - 7:00pm: Arrival 7:00pm - 7:15pm: Community announcements & short address by Emmanuel Malherbe, Director of Artefact Research Center 7:15pm - 8:00pm: Real-time collaborative editors in JupyterLab, Martin Renou 8:00 - 8:15pm: Handling imbalance data for client scoring, Abdoulaye Sakho 8:15 - 8:30pm: Choice modeling with the Python package choice-learn, Vincent Auriau 8:30pm - 10:00pm: Networking & drinks

Speakers Martin Renou Martin Renou is a Technical Director at QuantStack and a maintainer of Project Jupyter. Among other projects Martin is a core team member of the ipywidgets project and maintains many Jupyter widget packages such as ipyleaflet, ipydatagrid, ipygany, ipycanvas, and bqplot. He is a co-creator of the Voilà dashboarding system, and the xeus-python kernel.

Abdoulaye Sakho Abdoulaye Sakho is a graduate of ENSIIE and the M2 Random Modeling program at Université Paris Cité. He is currently doing a CIFRE thesis at LPSM (Sorbonne University) under the supervision of Erwan SCORNET and Emmanuel MALHERBE. He is a member of the Artefact Research Center.

Vincent Auriau Vincent is a PhD student at Artefact and CentraleSupélec, under the supervision of Vincent Mousseau. His work focuses on modeling customer preferences for assortment optimization. Previously he was a Data Scientist at L'Oréal R&I where he worked on different computer vision projects.

Abstracts Real-time collaborative editors in JupyterLab Jupyter has long been a powerful tool for interactive computing, but real-time collaboration has been a missing piece. In this talk, I’ll present our work on bringing true collaborative editing to JupyterLab, allowing multiple users to work on the same notebook simultaneously. The highlight of the talk will be a live demo showcasing JupyterGIS for collaborative GIS editing and JupyterCAD for real-time 3D CAD collaboration. I’ll also cover the technical foundations and challenges we faced in making these advanced editors work seamlessly in a collaborative environment.

Handling imbalance data for client scoring In data science, we talk about unbalanced data in a binary classification context when one class is under-represented compared with the other. Typically, it is the class of interest that is under-represented (cheaters, customers who bought a product on a site, sick patients...). In such a context, it is hard to train ML models. To tackle this challenge, various rebalancing strategies have been introduced in the literature. Throughout our presentation, we will focus on the oversampling strategies, which aim to generate synthetic data within the minority class to rebalance the data. SMOTE is the most common oversampling strategy, and many variants are based on it. We will also propose an alternative based on kernel estimation and generalized random forests. Finally, we will take a look at these techniques and observe their interest in a concrete use case from the banking industry.

Choice modeling with the Python package choice-learn Discrete choice models aim at predicting choices made by individuals from a set of options, called an assortment. Such models are used as inputs to operational problems such as assortment optimization, pricing or product recommendation. We present Choice-Learn, a Python package for choice modeling practitioners and researchers. The package enables processing choice data, and then formulating, estimating, and operationalizing choice models based on the TensorFlow library. We provide a unified implementation of classical choice models as well as neural network-based methods.

July is almost over and on 7 August we’d like to invite you to our meetup hosted by Artefact at their office in Utrecht, directly next to the Utrecht central train station! The meetup theme is centered around how to bring GenAI to production, but not just production, production to the next level. More specifically, the Artefact team will demonstrate how they use GenAI for the generation of customized, production-ready marketing images and the Phospho team will illustrate how to integrate robust ML best practices on scales of quality metrics for GenAI products. See you there!

SCHEDULE

• 18:00-19:00: Welcome with food and drinks! 🙌
• 19:00-19:45: Talk 1 - "GenAI Image Interaction: a next step beyond LLM text chatbots"
• 19:45-20:00: Break
• 20:00-20:45: Talk 2 - "Emerging best practices in Analyzing Usage Patterns and Quantifying Quality Metrics for GenAI Products"
• 20:45-22:00: Networking / drinks!

TALKS

[Talk 1]: “GenAI Image Interaction: a next step beyond LLM text chatbots" by Arthur Lambert & Priya Sarkar We've all experienced the capabilities of GenAI chatbots for data interaction. Now, it's time to explore the new GenAI innovations that Artefact is developing. Discover how GenAI is transforming marketing by accelerating asset creation and reducing costs. In this talk, we'll demonstrate how we're using GenAI to generate customized, production-ready marketing images. You'll also gain insights into automating the processes, enhancing efficiency for creativity based applications and learn about the quality metrics essential for monitoring and enhancing model performance.

[Talk 2]: “Emerging best practices in Analyzing Usage Patterns and Quantifying Quality Metrics for GenAI Products" by Paul-Louis Venard & Pierre-Louis Biojout Discover how to apply machine learning (ML) emerging best practices to Generative AI (GenAI) applications, specifically focusing on Large Language Models (LLMs) and diffusion models. This talk targets ML engineers and developers aiming to enhance their GenAI products through a quantified evaluation of model quality and user interaction analysis. Learn to implement rigorous, measurable standards to improve and understand GenAI applications.

The rapid advancement in Generative AI technologies, including LLMs and diffusion models, has empowered ML engineers and developers to build new and powerful products. However, the integration of robust ML best practices into the development of these products is still nascent. This session aims to bridge that gap by introducing established methodologies from traditional ML to enhance the reliability and effectiveness of GenAI applications.

DIRECTIONS Directly next to the Utrecht central train station, you’ll find the Artefact office in the Creative Valley building. The address is: Stationsplein 32, 3511 ED, Utrecht.

In this session, you will learn how to create a YAML Pipeline for Azure Data Factory in Azure DevOps for continuous integration and delivery. The pipeline validates the ARM template and exports the template into a build artefact that can be used by a release pipeline. The benefit of this process is that you can point your release pipeline to this artefact instead of the existing adf_publish branch.

Descripción de la sesión Conozcamos los nuevos recursos en el Centro de Administración de Power Platform nativo; como gestionar los servicios, complementos; asi como el monitoreo de Dataverse y demás artefactos de Power Platform.

Enlace Complementario https://learn.microsoft.com/power-platform/admin/admin-documentation/?wt.mc_id=3reg_22124_webpage_reactor

Orador Mayra Badillo Villamizar - Microsoft MVP

Évènement chez OCTO Technology (34 avenue de l'Opéra Paris)

80 places en physique - Merci de renseigner votre nom et prénom

#1 - CI/CD, la construction d’artefacts à l’heure du ML Sofia Calcagno - ML Engineer @OCTO Technology

La CI/CD permet de construire et déployer de manière automatique et robuste un logiciel. Le ML apporte un artefact supplémentaire à construire : le modèle de Machine Learning. Extrait du livre Culture MLOps, en 30 minutes, Sofia explorera les impacts sur la CI/CD de la gestion du modèle de ML.

#2 MLOps dans le secteur public : le cas des Finances publiques Claire Behar, Chief of Staff, équipe Data Engineering, Analytics & DataOps @DTNum au sein de la Direction Générale des Finances Publiques

De nombreux cas d’usages Data Science ont conduit les équipes de la Délégation à la transformation numérique (DTNum) de la Direction générale des Finances publiques (DGFiP) à mettre en place un prototype de ML Plateforme. Claire Behar partagera les challenges rencontrés ainsi que les perspectives de mise en place d'une démarche Green & MLOps, le numérique écoresponsable étant au cœur de la stratégie de l’État.

#3 REX: du Drift Monitoring dans l'aviation civile Mae Yener, MLOps Engineer @Hymaia

Un retour d'expérience à froid sur la conception d'un outil de monitoring du Model Drift pour Air France: enjeux, challenges, et ce que l'on peut en apprendre

Artificial Intelligence is at the core of Doctrine’s business. Over the years, they have developed over 30 Machine Learning models that are still running in production today. Two years ago, they realized that iterating over their models was becoming slow and cumbersome, mainly because of technical debt and a lack of automation. To solve this challenge, they assembled a task force of three engineers: David Huang, Ysé Wanono, and Aïmen Louafi. Over the span of a year and a half, they gathered the needs of Machine Learning Engineers, benchmarked tools, and then implemented a new architecture covering data labeling, exploration, hyperparameters tuning, experiment tracking, artefacts management, deployment, and monitoring.

⏰ Schedule

• 18h30 - 19h00: Welcome & Introduction 👋
• 19h00 - 19h45: Talk 🎙
• 19h45 - 20h00: Q&A Session 🙋‍♀️
• 20h00 - 22h00: Cocktail & Networking 🥂

🎙 About the speakers

David Huang is Machine Learning Engineer at Doctrine. During his career he has used Machine Learning on a range of topics: time series, images and, particularly at Doctrine, text. David is also a co-organizer of the Paris NLP meetups.

Ysé Wanono is Data Scientist and Data Engineer at Doctrine with previous experiences in statistical modeling at both consulting and tech startups. Ysé has been with Doctrine for more than 4 years, and for the past 2 years her focus has been on improving their MLOps lifecycle.

Aïmen Louafi is Machine Learning Engineer at Doctrine. He's fond of Computer Science, Mathematics, Machine learning. In the past, he had the opportunity to work with Graph Algorithms, OCR, Vector Search Engines, and much more.

🙌 Sponsors

Doctrine is a legal intelligence platform that uses technology to make legal analysis more credible.

nibble is the creator of spice, a Feature Platform for Machine Learning. We are also the organizers of these meetups :)

🤩 Audience

This is a meetup for data professionals who want to push forward the productionization of machine learning development within their organizations: data scientists, data engineers, software and devops engineers, data project/product managers, product designers, etc..

👋 Contact

For any inquiries, please contact [email protected]