
Smart equipment and continuous monitoring for water network
RESEARCH PROJECTS
Agentic AI, orchestration, governance, and compliance for EU defence
Through the MIDAS project, Almaviva is developing a modular, secure, and explainable AI middleware that integrates Large Language Models (LLMs) and intelligent agents to support decision-making across strategic, tactical, and operational levels (covering domains such as Strategic Command and Control - C2 -, C4ISR, logistics, and training), including within air-gapped, multilingual environments and fragmented legacy infrastructures.
The objective is not merely to introduce AI into defense systems, but to transform LLMs into reliable, transparent, and certifiable operational components embedded within real-world operator workflows.
Within the project, we contribute to the development of the system through two key dimensions: Agentic AI and AI Governance.
On the Agentic AI side, we develop adaptive and multimodal systems based on context-aware and role-aware architectures capable of supporting real-time operational decisions while ensuring that humans remain at the center of the command chain.
In terms of AI Governance, we develop an XAI framework focused on explainability, trustworthiness, and auditability, designed to comply with the requirements of the EU AI Act and GDPR in mission-critical environments.
Completing the overall architecture, the system’s LLM backbone also includes Velvet, Almaviva Group’s proprietary language model. Adapted to military terminology across EU languages, Velvet aligns with ISO 42001 principles and is designed for offline and air-gapped deployments.
reducing dependence on non-EU technologies and strengthening the European Defense Technological and Industrial Base (EDTIB)
by more than 15% and integration in 5+ military applications
WER below 5%, fidelity above 95%, and dialogue accuracy above 90%