DeepThink: Aligning Language Models with Domain-Specific User Intents