What data do we actually need for ML underwriting?

I’m researching this for a project and would love real-world input. If you’ve built or shipped something here, what did you learn? Context: I’m Jason Miller and I’m evaluating options for our team.