📄 Abstract
Digital wellness applications increasingly rely on large language models (LLMs) for personalised guidance, yet dependence on remote services introduces significant reliability risks. This study presents Hypnara, a full-stack digital wellness application using a resilient hybrid-AI architecture that integrates a cloud LLM with a local deterministic rule engine. Following design science research, the system was evaluated through specification-based verification and controlled fault-injection experiments. The architecture incorporates bounded behavioural context, a unified response contract, deterministic rule prioritisation, and allow-list prompt projection to restrict externally transmitted data. Under injected HTTP errors and service delays, the cloud-only configuration achieved valid-response rates of 90%, 70%, and 50% at fault rates of 10%, 30%, and 50%, respectively, whereas the hybrid configuration maintained 100% valid responses. When provider delays exceeded the one-second timeout, the cloud-only path failed completely, while the hybrid path preserved service through fallback, with p95 latency near the timeout threshold. Context-window testing further showed that shorter histories substantially reduced payload size. The design therefore supports graceful degradation during external AI service disruptions. Overall, Hypnara demonstrates that hybrid AI improves service resilience by preserving deterministic support while retaining cloud-based personalisation when available.
🏷️ Keywords
📚 How to Cite:
Thi Hien PHAN, Viet Tien LE , A RESILIENT HYBRID-AI ARCHITECTURE FOR PERSONALISED DIGITAL WELLNESS: DESIGN AND EVALUATION OF HYPNARA , Volume 11 , Issue 10, October 2026, EPRA International Journal of Research & Development (IJRD) , Pages: 91 - 97 , DOI: https://doi.org/10.36713/epra31823