A structured backbone for the messy, iterative reality of ML model development. Exploration and production deployment are treated as parts of the same process -- not separate phases -- reflecting how data science actually works: you explore, decide, build, evaluate, learn something new, and circle back.

Handles both entity-level and event-level datasets. Experiments and production can share the same tables without copying data (Delta Lake), features are served consistently across training and inference (Feast / Feature Store), and every experiment is tracked and reproducible (MLflow). Runs locally or deploys to Databricks.

Most ML tutorials jump straight to model.fit(). Real projects fail earlier -- in data issues you didn't notice, leakage you didn't check for, or feature choices you can't explain to your stakeholders three months later. This framework tries to close that gap.

It serves two audiences:

1. If you're learning, the notebooks walk through a realistic end-to-end process and explain the reasoning behind each step. Why does a 93-day median inter-event gap rule out short aggregation windows? Why might the model that wins validation degrade in production? The goal is to build intuition for the decisions that don't appear in textbooks.

2. If you're experienced, you can pip install, point to a new dataset, and get an opinionated exploration scaffold. The output is loosely-coupled production code (Bronze / Silver / Gold) with the provenance of every decision captured in self-contained HTML documentation -- useful when you need to explain why the pipeline does what it does.

• Exploration is a first-class concept. The framework records what it found in the data, what it recommends, and why -- in versioned YAML artifacts. Each downstream transformation traces back to a specific observation in a specific notebook, so nothing happens without a documented reason.
• Experimentation is version-controlled end to end. Not just code and features, but the actual data observations and actions taken on them can be frozen in time together. Delta tables support time-travel on live production datasets, so you can always go back to what the data looked like when a decision was made.
• Iteration is the default. Model feedback -- feature importances, error analysis, drift signals -- feeds back into the next exploration cycle. The framework tracks iteration lineage rather than treating each experiment as independent.

pip install "churnkit[ml]"

For Databricks, see the Databricks Installation guide.

churnkit-init --output ./my_project
cd my_project

Open exploration_notebooks/01_data_discovery.ipynb and set the data path:

DATA_PATH = "experiments/data/your_file.csv"   # csv, parquet, or delta

Execute cells sequentially. The framework auto-detects column types, data granularity (entity vs event-level), text columns, and temporal patterns -- then routes you through the relevant notebooks.

Findings, recommendations, and production pipeline specs are generated as you go.

Detailed documentation lives in the Wiki:

See CONTRIBUTING.md for guidelines.

Apache 2.0 -- See LICENSE for details.