Along this line of reasoning I learned to love Pydantic which can make it a breeze to parse and coerce environment variables to the correct types: https://pydantic-docs.helpmanual.io/usage/settings/

You can make an env.py and override module __getattr__ and then import environment variables just like they're regular Python objects (even booleans, floats, collections, etc, despite .env files being string only

Huge force multiplier for ML cuz then you can do hyperparameter optimization just by passing different environment variables in an outer loop (even inside your infrastructure as code)

Edit: you can make these classes immutable too: https://pydantic-docs.helpmanual.io/usage/models/#faux-immut...