ML explainability with a web UI your manager might actually use

What it does Shapash is a Python library that generates human-readable visualizations and interpretations of supervised models. It offers a web application to navigate between global and local explainability, filter predictions by features or errors, and inspect individual samples through an “identity card” summary. The library also compiles standalone HTML reports intended for model auditing and data governance.

The interesting bit The project treats interpretability as a communication problem, not just a technical one. It emphasizes explicit labels for every feature and value, and it evaluates the quality of its own explanations through stability, consistency, and compacity metrics—essentially auditing the auditor.

Key highlights

• Ships with a webapp for navigating global and local feature importance, with a live demo available.
• Generates standalone HTML reports for model auditing and data governance.
• Evaluates explanation quality with built-in metrics: Stability, Consistency, and Compacity.
• Supports most popular gradient-boosted and sklearn models: CatBoost, XGBoost, LightGBM, ensembles, linear models, and SVM.
• Allows feature grouping and custom color palettes to keep large models legible.

Caveats

• The README notes that unsupported models can be integrated, but it is vague about how much custom backend work is required.
• Much of the value is in the presentation layer; if you only need raw explanation tensors programmatically, this adds UI overhead you may not want.

Verdict Data scientists in regulated industries—or anyone who must defend model logic to non-technical stakeholders—should look here. If you just need programmatic explanation values in a notebook, it is probably overkill.