You explained a dynamic business‑logic model node for role management in your recommendation engine. How do you implement and manage these ORM and data models in Python?

I model roles as a separate table with a many‑to‑many relationship to users and a JSON column for dynamic permissions. Using SQLAlchemy, I create a Role class with a permissions JSON field and a UserRole association table. For dynamic business logic, I load role definitions from a YAML or database table and instantiate a RoleEngine that evaluates permissions at runtime. I keep the engine stateless and cache compiled permission trees in Redis, invalidating on role changes via a Pub/Sub channel. The repository pattern abstracts CRUD operations, so the recommendation service can query UserRole without knowing the underlying schema. I also use dependency injection (FastAPI’s Depends) to inject the role engine into request handlers, allowing unit tests to mock permissions. Periodic background jobs sync role definitions from an external admin UI, and I use Alembic migrations to evolve the schema safely. This approach keeps the role logic flexible, testable, and performant.