The Intelligent Backend for Time Matters

Sammanfattning

This thesis investigates the design and evaluation of an intelligent backend system for event-driven Just-in-Time Adaptive Interventions (JITAIs). The study addresses a key methodological challenge: how to systematically evaluate hybrid reasoning pipelines under real-time constraints in the absence of deployable real-world data. A simulation-based prototype is developed following a Design Science Research approach. The system integrates a controlled event generator, a complex event processing (CEP) layer for semantic transformation, and a hybrid reasoning module combining rule-based logic with large language models (LLMs). A closed-loop architecture is implemented, enabling intervention delivery to a VR client and feedback propagation into the system. The evaluation is conducted under controlled conditions across multiple dimensions, including performance, transformation behaviour, reasoning behaviour, correctness, consistency, and traceability. The results indicate that semantic event detection achieves stable segment-level performance under controlled replay conditions, with most errors arising from boundary effects in temporal aggregation. The system maintains stable decisionlevel outputs across repeated runs, including LLM-based reasoning under constrained structured output schemas. Latency analysis indicates that the reasoning stage is the dominant bottleneck, while upstream components scale efficiently with increasing load. A model comparison experiment suggests that a smaller LLM reduces latency by approximately 44% while preserving decision-level agreement, although it produces more conservative intermediate outputs. Overall, the findings indicates that simulation-based evaluation provides a viable and reproducible approach for analysing event-driven hybrid reasoning backends prior to deployment. The proposed architecture supports reliable semantic transformation, predictable latency, stable decision-level behaviour, and traceable decision pathways in controlled environments.