NutriSteppe-AI: Development, Architecture, and Explainable Design of a Large Language Model–Driven Chatbot for Personalized Health Menu Generation

Background/Objectives: Suboptimal dietary patterns are among the leading modifiable contributors to global morbidity and mortality, particularly in cardiovascular disease, type 2 diabetes mellitus (T2DM), obesity, metabolic syndrome, and hypertension. Digital nutrition platforms have emerged to improve adherence to evidence-based dietary strategies; however, many systems lack structured optimization, processing-aware nutrient profiling, and explainable artificial intelligence (AI) mechanisms. The integration of large language models (LLMs) into digital health introduces conversational personalization but also risks hallucination and unsafe outputs without constraint enforcement. This study aimed to describe the system development, architecture, database infrastructure, optimization algorithms, explainability enforcement, and digital health implications of NutriSteppe-AI, a chatbot-first LLM-driven system for personalized health menu generation constrained by deterministic nutrient logic and processing-aware scoring. Methods: NutriSteppe-AI integrates: (1) a multi-source structured nutrient database of 20,000 food products with up to 130 tracked nutrients; (2) energy requirement estimation using the revised Harris-Benedict equation; (3) linear programming-based multi-objective optimization; (4) a Healthy Food Index (HFI; 0.5–5.0 scale) incorporating NOVA processing classification penalties; (5) traffic-light nutrient gating; and (6) a constrained LLM orchestration layer governed by structured API contracts. Algorithmic validation was performed using 10,000 simulated user profiles spanning diverse age, anthropometric, activity, dietary exclusion, and budget parameters. Results: The system achieved 96.8% full constraint satisfaction with macronutrient mean absolute errors of 11.60% (energy), 18.86% (protein), 16.26% (fat), and 20.91% (carbohydrates). Incorporating NOVA processing penalties reduced ultra-processed food HFI scores by 0.73 points (P < 0.001). Median optimized menu HFI improved from 3.6 to 4.3. Median system latency was 1.8 s. Explainability validation confirmed 100% deterministic alignment with zero hallucinated numeric claims. Conclusions: NutriSteppe-AI demonstrates that LLM-driven nutrition chatbots can achieve deterministic, explainable, and clinically aligned performance when governed by structured optimization, processing-aware scoring, and explainability enforcement. This architecture provides scalable digital health infrastructure for cardiometabolic disease prevention in diverse populations.