As I was building models to make predictions on complex data, I struggled to evaluate how well they would generalize to new cases. Simple validation accuracy scores didn’t provide enough insight into real-world performance. So I decided to use k-means clustering for more unbiased model testing. By clustering my holdout test data into distinct groups of similar examples, I created an impartial scorecard for evaluation. I passed each cluster through my candidate models one by one. Analyzing performance per cluster revealed variability and biases that overall accuracy alone masked. Stepping back, it struck me how easy it was to hide gaps in model capabilities without profiling evaluation data. K-means clustering enabled me to ask precise questions to interrogate and improve model robustness. My next goal is to bring interpretability tools to my process for trustworthy predictive systems.