It’s a well-known statistic in the field: the vast majority of AI projects fail. But the most discouraging failures, with the greatest business impact, are those that occur once the model is in production: a poor recommendation model erodes user trust in the algorithm, a faulty facial recognition model prevents phone unlocking, a bad pedestrian detection model can cause a fatal accident…

Over the past four years, I have deployed several AI algorithms in different contexts: each presented challenges and learnings that have shaped my beliefs on how to properly deploy AI algorithms (MLOps). I will share the most interesting insights from these experiences in this talk :

1) The importance of load testing

2) Watch out for data drift!

3) How to know if my model works in real life?

4) Let’s (re)discover together through these experiences the fundamentals of ML monitoring!