Fifteen years ago, teams were struggling to deliver MATLAB models to engineering teams because of real technical challenges (porting to C/C++, floating-point vs fixed-point computations, assembly optimizations, etc). In 2026, the machine learning ecosystem has mostly converged to Python for the entire lifecycle with a promise to break down silos and streamline the deliveries from research to engineering to production. But in reality, data science and engineering teams still struggle to deliver smoothly.

What if the core issue was not technical but human? How to make data science care about engineering, and vice-versa?

In this talk, we'll cover some common situations from the field of MLOps, among which:

• Start simple and iterate
• Resist the OOP fever
• Deploy your notebooks
• Deliver models using standard formats
• Prioritize real-world usage and data

You will leave with some pragmatic actions, illustrated with open-source technologies mostly in the Python ecosystem (such as Papermill, Marimo, ONNX, MLflow).

This light-hearted talk is intended for data scientists, ML/Data/Software engineers and technical leads looking for a pragmatic angle to reduce the time to delivery of ML-driven products.

Some familiarity with data science activities and/or software engineering is advised.