You don’t need to be expert in every ML algorithm. But you need to know at a high level what types of problem each algorithm is better suited to solve.
Do we have labeled data? Do we have a lot of labeled data? If you have a lot of labeled data, use supervised learning. If you have some but not a lot of labeled data, use semi-supervised learning. If you don’t have labeled data, use unsupervised learning, or consider how to get data labeling first.
Understand correlation is not the same as causality. This is key in feature selection. Having clouds in the sky and people carrying umbrellas are both events that have a high correlation with rainy weather, but if you are building a model to predict rain, you can't use umbrella as the feature.
Watch out for data distribution changes. Whenever the data distribution changes, the model needs to be rebuilt/retrained. Therefore, it's critical to monitor when such changes occur.
Understand most of the real work is in data cleaning. In the classroom, we are always given a copy of cleaned data and begin building a model from that data. In reality, data is never clean and >90% of the work is spent on cleaning the data so that we can even start modeling.