Lessons learned from deploying Machine Learning in an old-fashioned heavy industry

Simple machine learning models (linear regression, random forest) often outperform complex ones for industrial applications. Avoid spending too much time on complex models.

Customer data is extremely challenging to work with - issues include data quality, calibration changes, and different standards between facilities.

Machine Learning as a Service (MLaaS) is much harder to scale than Software as a Service (SaaS) - requires more consulting, customer support and domain expertise.

Proper model evaluation is critical - use temporal cross-validation instead of random cross-validation to respect time-based relationships in the data.

Most of the work is infrastructure, not ML - data pipelines, monitoring, configuration management, and customer-facing interfaces are larger parts than the actual models.

Domain expertise and ability to communicate with customers in their language is essential - customer success personnel are often more important than ML engineers.

Regular retraining and monitoring of models is necessary due to machine recalibrations, seasonal changes, and other real-world factors.

Aim to control your own data collection rather than relying on customer data when possible to ensure quality and consistency.

Simple, interpretable models help gain customer trust and make troubleshooting easier in industrial settings.

Real-world industrial applications often have significant delays between predictions and ground truth (like 28-day cement strength tests), which must be accounted for in the ML system design.