Machine learning models often exhibit inconsistent performance across different parts of a dataset, making summary metrics like AUC and F1 insufficient for identifying areas needing improvement. Tools like Uber’s Manifold and TensorFlow Model Analysis (TFMA) enable more detailed visualization of model performance across data slices. In this context, the TFMA library is utilized to visualize evaluation metrics variation across different dataset feature slices and subsequently log these visualizations to Comet for easy comparative analysis of different models. The process involves setting up TFMA by downloading necessary datasets and schema, converting data to TFRecords format, and defining evaluation configurations using protobufs for model analysis. Results are then logged to Comet, allowing visualization through TFMA Viewer Custom Panels, which include various plots such as Residuals, Calibration, Precision-Recall, and ROC curves. These visualizations aid in debugging models by comparing performance against target values and determining cutoff thresholds for classifiers, especially on imbalanced datasets. Additionally, TFMA panels at both project and experiment levels facilitate analysis across multiple models or within a single model, and offer a feature to run a diff of two experiments, enhancing model evaluation and comparison capabilities.