Working on building a monitoring and log trace stack for machine learning. What’s weird/hard is that we’re looking to deduce the performance of models that do not necessarily have ground truth readily available so it’s tricky to figure out if the model is working or not.