Engineering ML & AI at Scale: Feature Selection, Explainability & Agentic Engineering

Calling all data scientists, machine learning engineers, and data enthusiasts! Join us on Wednesday, Oct 28 at Mollie. We're bringing together engineers from Mollie and Bonnie to share hard-earned lessons from the trenches—covering everything from feature selection and interpretable AI predictions to practical methods for navigating AI-generated code to help teams ship better software. Stick around after the talks for good food, drinks, and great conversations with fellow tech enthusiasts. 📋 Agenda 17:30 - 18:25: Welcome with food and drinks! 🍕🍺 18:25 - 18:30: Mollie’s Intro 18:30 - 19:00: Talk 1: Feature Selection in Production: Lessons from Fraud Detection, by Nicole Van de Weijer (Mollie) 19:00 - 19:35: Talk 2: Machine Learning Model Explainability in the AI Age, by Tim Haarman (Mollie) 19:35 - 19:45: Break ☕ 19:45 - 20:15: Talk 3: Reviewing the reviewer: bring rigour and discipline to your code base in the AI (slop) era, by Anne Lohmeijer (Bonnie) 20:15 - 21:00: Networking & Drinks 🎤 Talk Details Talk 1: Feature Selection in Production: Lessons from Fraud Detection, by Nicole Van de Weijer Managing engineered features in a production fraud model means trading off accuracy against latency and interpretability. In this talk, I'll walk through the feature selection options we considered for Mollie's consumer fraud detection model, comparing manual selection to automated methods, Backward Elimination (RFECV with SHAP or permutation importance) and Forward Selection (MRMR / MRMRCV), and their trade-offs. I'll close with the method we settled on and why. Talk 2: Machine Learning Model Explainability in the AI Age, by Tim Haarman Building an accurate model is only half the job: if users don't trust its predictions, they won't act on them. Understanding why a model made its choice is key to that trust. This is nothing new - we’ve had SHAP for years as a standard tool for explaining predictions - but it speaks the language of data scientists. Telling an analyst a feature has a SHAP value of +1.40 builds little confidence. We show how we combine SHAP with AI to turn technical attributions into grounded, human-readable explanations analysts can act on immediately. Talk 3: Reviewing the reviewer: bring rigour and discipline to your code base in the AI (slop) era, by Anne Lohmeijer Nowadays, generating code is easier than ever. Agents make fewer mistakes, and merge requests are getting bigger. Does that mean the code is also getting better? The question of how to adapt as a software engineer to this rapidly changing set of tools is coming back more and more. It's one of the challenges I have been trying to tackle since joining Bonnie a year ago. I’m going to open-source my experience and share the different approaches I’ve taken over the last year to bring discipline and rigour to the way a fast-paced, engineering team ships features in an AI startup. Direction Mollie Amsterdam Office 📍Address: Keizersgracht 126, 1015 CW, Amsterdam Mollie’s office is right in Amsterdam’s historic city center. It’s a nice ~20-minute walk (about 1.5 km) from Central Station. If you’d rather hop on public transport, trams 2, 17, or 24 will get you from Central Station to Keizersgracht in ~10 minutes.