Knowledge-Driven Hybrid Architecture for Multi-Modal Decision Support : A Case Study in Renovation Offer Generation

Sammanfattning

Complex service domains, such as renovation planning, require structured decision-making processes where consistency, traceability, and correctness are critical. At the same time, recent advances in artificial intelligence (AI), particularly Large Language Models (LLMs), enable more flexible and natural interaction with users, but introduce uncertainty due to their probabilistic nature. This creates a fundamental design challenge in decision support systems (DSS): how to combine the flexibility of AI with the reliability of deterministic logic. This thesis addresses this challenge through the design and implementation of a knowledge-driven hybrid architecture for decision support. The proposed approach separates probabilistic and deterministic responsibilities by confining AI components to interaction and interpretation, while a deterministic decision engine governs state management, constraint enforcement, and consequence evaluation. Domain expertise is encoded as a graph-based decision schema with guarded activation and forward propagation, enabling explicit modeling of interdependencies between decisions and their resulting actions. The system is developed and evaluated in collaboration with a renovation company, Dryft, where it is applied to the generation of structured renovation offers. The evaluation follows a design research approach and assesses the system with respect to correctness, consistency, and expressiveness through representative case scenarios. The results demonstrate that the proposed framework produces consistent and traceable outputs across multiple interaction modalities, while maintaining the flexibility required for handling user input and real-world complexity. The main contributions of this thesis are a formal decision modeling framework for structured service configurations, an architectural design principle for hybrid AI systems, and an empirical case study demonstrating the applicability of the approach in a real-world domain.