The recent advancements in Large Language Models (LLMs) have demonstrated unparalleled versatility in executing diverse tasks. Challenges arise when applying these models to user-facing systems. These models are prone to hallucination and often lack explainability, making them unreliable for critical Decision making in specialized domains such as nutrition. The need for real-time expert systems that enable reliable decision-making is evident. Furthermore, ensuring fairness and reducing biases AI-driven systems remains a significant challenge. To address these limitations, we introduce NutriChat, a multi-agent architecture designed to facilitate knowledge sharing between agents and enhance reasoning capabilities. By leveraging a debate- driven approach enriched by the HUMMUS dataset, NutriChat integrates Human-in-the-Loop mechanisms to improve interpretability and personalization in decision making, resulting in an ability to chart meal plans at 0.67 efficacy and a 0.90 adherence to user preference. This novel approach advances agentic systems to prioritize the user’s preferences thereby reducing harmful bias.