An sklearn-compatible Python toolkit bundling 14 hybrid tabular classifiers, experiment protocols, fairness audits, drift monitoring, and a structured HTML report generator behind one drop-in package.
clinikit is a lightweight, sklearn-compatible Python toolkit for classification on tabular data. It started from a simple experiment goal: building a more defensible, controlled, and reproducible workflow around the PIMA diabetes dataset.
The process led to a package that brings together 14 hybrid classifiers, five experiment protocols, calibration and label-noise diagnostics, subgroup fairness auditing, drift monitoring, and structured HTML report generation. clinikit is not designed for production deployment. It is designed for research and development workflows.
The tabular ML ecosystem is powerful, but fragmented. A typical applied workflow often needs scikit-learn for core estimators, Cleanlab for label-noise checks, Fairlearn for subgroup analysis, MAPIE for conformal prediction, and manually written scripts for reporting.
Each tool has its own API, configuration style, and reproducibility story. When the goal is to build a quick baseline or produce a defensible experiment report, connecting these tools can take longer than the modeling itself. clinikit was created to reduce that fragmentation.
clinikit provides 27 modules through a single import surface:
check_estimator tests.clinikit train | benchmark | audit | validate | report.The goal is not just to produce a single score. The goal is to make the experiment process more readable, auditable, and reproducible.
clinikit was released on PyPI in May 2026 as v0.1.0, followed by 0.1.1 with small improvements. The package is MIT licensed and supports Python 3.10, 3.11, 3.12, and 3.13. It is tested with CI on Ubuntu, macOS, and Windows.
The current version passes 485 tests and has approximately 89 percent line coverage. Installation takes one command: pip install clinikit.
The API design, implementation of 27 modules, test suite, CI/CD pipeline, and first PyPI release were built end to end. What started as a single experiment goal became an open-source toolkit for researchers and developers working with tabular data.