« All happy families are alike; each unhappy family is unhappy in its own way. »
Tolstoy's opening line may resonate when you have worked on machine learning models for a long time. Successful models tend to look alike: clean data, reliable data collection, hard work to get everything right, and enough stakeholder trust to hand over the reins in production. On the other hand, every failed model fails in its own way - a training set that silently drifted, a feature that leaked the future, a business constraint no one thought to verify, or subtle context shifts once the system went live. No two headstones read the same.
After 20 years of designing and deploying models in banking, insurance, e-commerce, and hedge funds, I have cataloged many of these failure modes firsthand, and I would like to share what they taught me. I trained in mathematics at École Polytechnique, co-founded a startup in Silicon Valley, managed a complex analytics infrastructure in San Francisco, consulted within McKinsey's Technology arm, and spent years managing data-heavy programs at Société Générale's Corporate and Investment Bank. In 2011, I founded ArsChitectura, a boutique consulting firm focused on machine learning for high-stakes industries. Each of these chapters left a fingerprint.
Along the way, I have implemented predictive models and AI-driven systems across these sectors - from credit scoring and insurance pricing to real-time recommendation engines and trading signals. That range of problems and failure modes revealed things that easy "do-it-all" approaches tend to overlook. The important things, it seems, only unfold with time.
That is what this blog is about. Not theory in isolation, but the bruised insights from applying machine learning where the stakes are real: where a model's output drives business decisions, where business managers need transparency on a model's behavior, and where failure has tangible consequences. The gap between an interesting algorithm and a reliable production system is vast, and it is in that gap that most business value resides.
In future posts, I will write about data riddles, model design, unusual feature engineering, explainability, and the pitfalls I have encountered along the way. Whether you are an executive or manager responsible for the projects that depend on these systems, or a data scientist or machine learning engineer building them, I hope you will find practical, honest reflections here. After all, you need to get lost first to draw a map.
If any of this resonates or you see things differently, I would enjoy hearing from you.