Embodied conversational agents and Large Language Models have revolutionized the way humans interact with machines. Because they are trained with vast amounts of knowledge, these technologies could pave the way towards a scalable, privacy-preserving, and more inclusive way of offering services to vulnerable and stigmatized communities. In this work, we adapt state-of-the art models GPT-4o and DeepSeek-R1 to a dataset of request from autistic users, using a retrospective feedback loop and constraints designed to boost empathy and favor user-centered preferences. We maximize user satisfaction by incorporating a module of sentiment analysis with hierarchical attention networks. Our system achieves outstanding performance compared with traditional machine learning baselines and human responses. However, a multimodal approach taking into account facial expressions and voice signals may further enhance the performance. 

- *Written by Clémentine Bleuze*