Notes pratiques issues de la recherche.
Des essais de fond sur ce qui fonctionne réellement lorsque l’IA passe en production — et sur les défaillances discrètes.
Dernières publications
Chaque essai examine la mise en œuvre concrète du MLOps, des prévisions, de la vision par ordinateur, du traitement du langage, de la recherche documentaire et de la gouvernance. Les articles complets sont actuellement rédigés en anglais.
The future of disaster response: AI's impact on emergency relief
Where AI genuinely helps emergency relief — damage mapping, demand forecasting, multilingual triage — and where human judgement must stay in charge.
Five MLOps mistakes we keep seeing in 2025
From feature-store sprawl to dashboards no one looks at — what we'd do differently after a decade of shipping models into production.
Why your demand forecast is biased (and what to do about it)
A practical playbook for spotting and removing systematic forecast bias — drawn from a decade of forecasting engagements.
Choosing between classical CV and deep learning
When OpenCV beats a neural net — and how to know in advance, before you have invested six months in the wrong stack.
Aspect-based sentiment without breaking the bank
A small-model recipe for accurate, fast, multilingual aspect sentiment that does not require a frontier LLM in the hot path.
Build vs buy in 2025: the new rules
How the rise of foundation models has rewritten the build-vs-buy calculus — and the rules that have stopped being true.
Reproducibility patterns we actually use
The minimum viable practices that make experiments trustworthy — drawn from systems we have been running since 2014.
RAG in the enterprise: what works in 2025
Retrieval-augmented generation has matured. Here is the architecture we now reach for by default — and the failure modes we still see weekly.
The AI Act, for practitioners who have to ship anyway
What the EU AI Act actually requires of high-risk and limited-risk AI systems — and the documentation we have been writing on every project since 2023.
The quiet comeback of small models
The 7-billion-parameter and smaller models we deploy in 2025, the gaps they fill, and why the cost story alone is not the most interesting one.
Un problème mérite-t-il un article ?
Parlez-nous de la difficulté rencontrée par votre équipe — les meilleurs essais partent de questions concrètes.