Transforming informal stakeholder descriptions into clear and actionable software requirements is a critical challenge in requirements engineering. Existing Artificial Intelligence (AI)- based tools can process natural language but often overlook alignment with underlying business processes and coverage of all relevant scenarios. This paper proposes a hybrid neuralsymbolic framework, referred to as Process-Aware Requirement Generation (PARG), which integrates transformer-based process recognition with ontology-guided requirement generation. PARG extracts key process elements, such as actors, actions, conditions, and outcomes from user stories, maps them to domain-specific processes, and instantiates structured requirement statements using predefined templates. The framework further validates consistency and evaluates coverage to ensure completeness and traceability. By combining language understanding with structured process knowledge, PARG reduces manual effort, supports improved traceability, and produces requirements that are coherent, process-aware, and aligned with real-world workflows.

Resilient digital transformation increasingly depends on trustworthy cyber-physical software, where Digital Twin (DT) ecosystems—virtual replicas of low-power Internet of Things (IoT) devices, edge gateways and industrial assets— must remain available, observable and explainable under attack. These environments are particularly exposed to zero-day threats that exploit the semantic gap between physical states and virtual expectations, while heterogeneous topologies, concept drift and constrained edge resources erode the assumptions of conventional Intrusion Detection Systems (IDSs). We present Ψ-Risk-DT, a Neurosymbolic (NeSy) framework that couples entropy-based anomaly detection with an Associated Random Neural Network (ARNN) and an RDF/SPARQL semantic-reasoning layer through a formal entropy-gated operator Ψ; a Modular Semantic Update (MSU) mechanism rewrites only the affected portions of the DT knowledge graph, and a hybrid loss aligns classification accuracy, graph coherence and semantic consistency with Lyapunov-style stability guarantees. The framework is directly validated in a containerised Network Time Protocol (NTP) amplification scenario representative of volumetric zero-day-like attacks against DT ecosystems, where it achieves an Area Under the Curve (AUC) of 0.993, a False-Positive Rate (FPR) of 0.021 and an end-to-end pipeline latency of approximately 25.8 ms, while MSU reduces the symbolic-update load by up to 88% relative to a global-rewrite ablation evaluated on the same trace. For comparative positioning, these figures are contrasted with values reported in the literature for established IoT/DT IDS baselines (Kitsune, deep-learning zero-day detectors, NeSy-IDS and the (H-DIR)2 predecessor); under the conditions reported in those works, Ψ-Risk-DT exhibits an order-of-magnitude latency gain[C] and a 20–40% FPR reduction[C] relative to deep-learning baselines, presented as a positional comparison rather than as a coevaluated measurement. Broader empirical replication on out-of-distribution (OOD) variants of community benchmarks (Kitsune, WiseML 2024, Sec4ML 2023), on multi-vector Routing Protocol for Low-Power and Lossy Networks (RPL) attack scenarios, and on heterogeneous edge hardware is identified as ongoing future work. Overall, Ψ-Risk-DT provides explainable, entropy-gated, semantically adaptive protection for DT ecosystems and contributes to FSOFT topics on resilient digital transformation, trustworthy cyber-physical software, IoT security, adaptive threat detection and explainable neurosymbolic Artificial Intelligence (AI).

Domain-Specific Languages, apart from the design and formal definition aspects involved to build, also require Model-to-Text (M2T) transformations to generate executable code. Authoring the M2T templates is a manual, labor-intensive process that often generates frustrating, time consuming errors. And, although Large Language Models have advanced code and grammar generation, the automated synthesis of downstream Model-to-Text templates still remains largely unaddressed. In the context of software engineering and model-driven development, automating such transformations addresses a core challenge in modern software toolchain construction. This paper presents a work-in-progress multi-agent system that automatically generates production-ready Jinja2 templates from a textX grammar, natural language requirements and sample model instances. The architecture decomposes the generation task into seven stages orchestrated by eleven specialized agents, employing deterministic runtime introspection to provide ground-truth structural data. To ensure reliability, the generation process is coupled with a rigorous three-level deterministic validation loop that evaluates template syntax, rendering execution and target-language correctness without relying on self-assessment by the language models. We demonstrate the pipeline on two structurally versatile domains, specifically a recursive calculator and a declarative smart home automation language, illustrating its ability to produce functional templates, rendered code and test artifacts across diverse grammar paradigms.

The evolution of software engineering is increasingly driven by the integration of artificial intelligence and digital paradigms. In critical domains such as medical and biological imaging, modern software systems are transitioning from static, rule-based workflows to dynamic, AI-powered environments. This paper presents the design, implementation, and preliminary validation of an advanced, fully automated software pipeline for Transmission Electron Microscopy image analysis to address software engineering and analytical challenges in tendon fibril segmentation. The proposed architecture replaces a previous hybrid segmentation approach with a deep learning model based on the ViT-UNet architecture, integrating targeted preprocessing modules, advanced segmentation capable of capturing both local details and global spatial relationships, and an automated feature extraction and clustering engine. Initial experimental results show an improvement in clustering accuracy from 57.14% to 64.71%. While these findings suggest the potential of Transformer-based architectures to enhance medical image analysis, they also highlight the need for further evaluation on larger datasets to establish broader statistical significance and validate their overall effectiveness.