Edge AI-based systems distribute intelligence across hetero- geneous, resource-constrained environments along the continuum. De- spite growing adoption, their architectural practices remain poorly con- solidated in open-source software, and the label Edge AI is often ap- plied to artifacts with uneven architectural relevance. This paper re- ports an empirical study of open-source GitHub repositories identified through repository mining and analyzed through qualitative coding in- formed by ISO/IEC 30141, ISO/IEC/IEEE 42010, and ISO/IEC 25010. We examine capability distributions, architectural patterns, documenta- tion practices, and architecturally relevant quality concerns. The results show a predominance of supporting, interface, and data capabilities, with most systems emphasizing infrastructure-oriented and cross-layer con- cerns across device, edge, fog, and cloud environments. Architectural knowledge is documented mainly through executable and operational artifacts, such as configuration, interface, and deployment files, rather than explicit architecture descriptions. Quality concerns are strongly runtime-centered, especially reliability, maintainability, and performance efficiency, while concerns such as safety assurance and energy awareness appear only marginally. These findings expose recurring blind spots in open-source Edge AI practice and provide an empirical basis for future architectural guidance and evaluation.