LLM-empowered software systems for constrained environments

Modern software systems increasingly operate in constrained environments, where activities must be organized and coordinated in accordance with technological or business requirements whose violation may compromise correctness, efficiency, safety, or organizational goals. In technologically constrained environments, software services must be scheduled, deployed, and coordinated while respecting timing and resource constraints. In business-constrained environments, process activities must be analyzed through measurable indicators that reflect organizational objectives. In both cases, a critical step is the transformation of informal descriptions or data-based information into structured artifacts that can be validated and processed by software tools. This thesis investigates the controlled and structured use of Large Language Models (LLMs) to support this transformation. Rather than using LLMs as autonomous decision-makers, the proposed approaches embed them within pipelines that combine domain-specific context, intermediate formal representations, validation and repair mechanisms, and deterministic computation. The technological dimension is addressed through a constraint-based formulation of workload scheduling, the WS-Line pipeline for generating optimized and executable schedules from technical documentation, and a deployment framework for containerized real-time microservices. The business dimension is addressed through two complementary approaches: PPIPilot, which adopts a top-down strategy to suggest and compute Process Performance Indicators (PPIs) from process descriptions, goals, and event logs; and Declare2PPI, which adopts a bottom-up strategy to derive PPIs directly from event logs through declarative process mining.