Embedding DNNs in resource-constrained systems is gaining increasing interest across various domains. However, deploy- ment remains hindered by opaque, vendor-specific tools that limit transparency, flexibility, and reproducibility issues. In this paper, we introduce Aidge, an open-source framework for the design and deployment of DNNs in constrained environments. Aidge is built around four core principles: being community- driven and dependency-free, offering a user-friendly hierarchi- cal graph intermediate representation, enabling full traceability, and ensuring two-way interoperability with major deployment tools. These principles guide a modular architecture that al- lows users to analyze, optimize, and deploy models with a high degree of control and portability. By reducing reliance on black-box toolchains and supporting transparent, verifiable workflows, Aidge offers a practical and robust foundation for embedded AI development, particularly in applications requir- ing high assurance, such as safety-critical systems.
Moineau et al. (Tue,) studied this question.