What steps would you take to troubleshoot a model that is underperforming in a production ML pipeline?
