As Social-Emotional Learning (SEL) requires personalized, context-aware interactions, educators seek more adaptive teaching support. Generative AI addresses this by providing automated and personalized assistance. However, powerful AI models like Large Language Models (LLMs) still lack the capacity for emotionally nuanced responses grounded in real-life experiences and cultural subtleties, which are essential for SEL education. Therefore, we introduce beSEL (boosting educators for social-emotional learning), a teaching assistant system that aligns LLMs with human expertise in SEL through two core functionalities: (a) collaborates with educators to optimize lesson plans, and (b) engaging with students to support SEL knowledge and skill acquisition. We designed an efficient pipeline to construct domain-specific SEL instructional datasets via multi-round LLM-based segmentation and matching. Using these datasets, we fine-tuned Qwen2.5-3B using LoRA and evaluated the system to assess whether the generated content meets SEL teaching standards. We find that beSEL demonstrates more sufficient emotional intelligence and situational understanding than several general AI models, highlighting the effectiveness of SEL human-alignment by supervised fine-tuning.