Why Problem Framing, Not Algorithms, Is the Real Challenge in Machine Learning

A software engineer transitioning from SRE and DevOps to AI engineering argues that the hardest part of machine learning is not the mathematics but the upstream thinking that precedes any model. Drawing on a real-world example of predicting SLA breaches in IT services, the author outlines four critical questions that must be answered before training begins: what event is being predicted, what action follows the prediction, what a wrong answer costs, and where the model's uncertainty becomes dangerous. The author contends that algorithm selection — whether to use classification, regression, or other approaches — resolves itself naturally once these foundational questions are properly framed. This approach mirrors practices from site reliability engineering, where monitoring alerts are never deployed without clearly defining what triggers them, who responds, and what a false alarm costs. The central takeaway is that a well-framed problem makes the algorithm choice almost automatic, which is why it should be the last decision made, not the first.