The rapid advances in LLMs present tremendous opportunities to create interactive, personalized learning experiences at scale. However, to truly harness the educational potential of these technologies, it's crucial that teachers - who are at the forefront of daily student interaction and possess indispensable knowledge and expertise - go beyond being mere consumers to key contributors. This one-year project proposes a teacher-AI collaboration paradigm that can potentially realize this vision. In this project, we partner with WGBH Educational Foundation (GBH) to support teachers to create and customize intelligent tutor accompanying the STEM learning resources available on PBS LearningMedia. We have performed an interview study with high school science teachers to understand how they teach videos in their lessons. Teachers overwhelmingly liked engaging students in a conversation about the video as a method for inquiry and learning. The interview study shed light on the tutoring strategies teachers would use. For example, teachers make sure the questions do not go beyond the scope of the video, and they do not probe further when the students begin to get frustrated. Based on the findings, we are developing a platform that enables teachers to configure LLM-powered conversational tutors for STEM videos. We employ LLM alignment techniques and human-AI collaborative techniques to ensure that the conversational tutors provide a safe and desirable learning experience to students.