How would you evaluate and mitigate bias in a large language model deployed for a customer-facing application?