NutriSteppe-AI: Development, Architecture, and Explainable Design of a Large Language Model-Driven Chatbot for...

Area of Science:

• Digital Health and Nutrition Informatics
• Artificial Intelligence in Healthcare
• Computational Nutrition and Dietetics

Background:

• Suboptimal dietary patterns are major contributors to global morbidity and mortality, especially cardiovascular disease, type 2 diabetes mellitus (T2DM), obesity, metabolic syndrome, and hypertension.
• Existing digital nutrition platforms often lack structured optimization, processing-aware nutrient profiling, and explainable AI, limiting their effectiveness.
• Large language models (LLMs) offer personalized digital health solutions but require careful constraint enforcement to prevent hallucinations and ensure safety.

Purpose of the Study:

• To describe the development, architecture, and digital health implications of NutriSteppe-AI, a chatbot-first LLM-driven system for personalized health menu generation.
• To detail the system's constraints, including deterministic nutrient logic and processing-aware scoring, for safe and explainable AI outputs.
• To evaluate the system's performance in generating clinically aligned and optimized nutritional plans.

Main Methods:

• Integration of a structured nutrient database (20,000 foods, 130 nutrients), energy estimation (revised Harris-Benedict equation), and linear programming optimization.
• Implementation of a Healthy Food Index (HFI) with NOVA processing classification penalties and traffic-light nutrient gating.
• Development of a constrained LLM orchestration layer with structured API contracts, validated using 10,000 simulated user profiles.

Main Results:

• Achieved 96.8% full constraint satisfaction with acceptable macronutrient errors (11.60-20.91%).
• NOVA processing penalties significantly reduced ultra-processed food HFI scores (0.73 points, p < 0.001), improving median HFI from 3.6 to 4.3.
• Demonstrated 100% deterministic alignment with zero hallucinated numeric claims, with a median system latency of 1.8 seconds.

Conclusions:

• NutriSteppe-AI proves LLM-driven nutrition chatbots can achieve deterministic, explainable, and clinically aligned performance through structured optimization and constraint enforcement.
• The system architecture offers a scalable digital health infrastructure for personalized nutrition and cardiometabolic disease prevention.
• This approach addresses limitations of current digital platforms by integrating advanced AI with robust nutritional logic and safety mechanisms.