As artificial intelligence becomes more and more a part of education, the challenge is not about having access to generative tools, but about connecting them with the goals of the curriculum and the needs of the classroom. This chapter presents the design and evaluation of a large language model–based chatbot developed specifically for teaching quadrilaterals in lower secondary mathematics. The chatbot integrates fine-tuning with retrieval-augmented generation (RAG), combining accurate, curriculum-aligned content with flexible, conversational support. The chatbot allows learners to ask conceptual questions, solve problems step by step, receive guided hints, and generate flashcards or exercises of varying difficulty. A hybrid routing mechanism selects the most appropriate response strategy based on user intent. Evaluations using both isolated prompts and multi-turn dialogues demonstrate that the hybrid system significantly outperforms standard LLM baselines in terms of accuracy, consistency, and pedagogical suitability. A classroom trial with 20 students confirmed the tool’s usability and effectiveness; students reported high satisfaction and meaningful engagement. This study demonstrates that, with careful content and architectural structuring, generative AI can enhance student learning while supporting differentiated instruction. Future directions include scaling the approach to other topics and incorporating multimodal capabilities.