A Multi-Agent LLM Framework for Dynamic Constitutional (Prakriti) Profiling and Personalized Ayurvedic Interventions

Traditional Ayurvedic diagnostics rely on determining an individual's baseline constitution (Prakriti) and acute functional imbalances (Vikriti) through qualitative clinical evaluations and longitudinal lifestyle assessments. While foundational to personalized holistic care, manual constitutional profiling exhibits inter-practitioner variability, requires substantial clinical time, and lacks continuous adaptability to acute physiological or environmental shifts. This paper presents a Multi-Agent Large Language Model (LLM) and Rule-Hybrid Architecture designed to structure constitutional classification and generate source-grounded lifestyle (Vihara), dietary (Ahara), and botanical (Aushadha) educational recommendations. By decoupling diagnostic sub-tasks across specialized cooperative agents and enforcing a deterministic rule-verification engine, the proposed framework separates stable baseline traits from transient symptom fluctuations while constraining unsupported generative outputs. Evaluated on a held-out synthetic benchmark cohort of 200 multi-modal profiles (derived from a 250-profile curated dataset) labeled by a consensus panel of three certified Ayurvedic physicians, the framework achieves an overall classification accuracy of 90.0%, a macro-F1 score of 0.892, and a system-to-consensus Cohen's kappa of κ=0.863(95% CI: 0.808–0.918). Ablation experiments against deterministic rule baselines and monolithic single-LLM configurations demonstrate that multi-agent task separation combined with deterministic safety gating eliminates 100.0% of predefined contraindication rule violations across benchmark and adversarial stress tests. Adhering to TRIPOD+AI reporting standards, we detail the knowledge-base provenance, system failure modes, computational overhead, and human-in-the-loop governance required for responsible deployment in integrative health informatics.