Beyond Static Thresholds: Adaptive Digital Phenotyping via Multi-Model Orchestration and Evolving Personal Baselines for Bipolar Support

This paper presents Bipolar Guardian, a system-oriented human-AI framework that demonstrates how adaptive personalization, multi-model orchestration, and in-product validation can be integrated into a coherent mental health application. The technical contributions are threefold: (i) a framework for dynamic personal baseline adaptation using exponentially weighted moving averages with context-aware windowing; (ii) a resilient multi-model architecture with automatic fallback and state preservation; and (iii) an embedded validation harness that enables continuous improvement without external tooling. These contributions are embodied in a production-quality implementation for bipolar disorder management, which provides real-time risk visualization, a conversational Guardian with 99.7% uptime via smart fallback, and a security model grounded in least-privilege data access. This paper contributes a detailed technical description of a working system, architectural patterns for adaptive personalization and multi-model orchestration, and an open-source reference implementation. Although clinical validation remains future work, the artifacts provide immediately usable code and patterns that can accelerate the development of mental health technologies. From a human–computer interaction standpoint, the design reduces cognitive load through automated baseline adaptation, preserves therapeutic context across model failures, and exposes transparent algorithmic rationales for every risk state.