Small Language Models (SLMs) are increasingly being deployed on mobile and edge devices, offering benefits such as reduced infrastructure costs, improved privacy, and offline access. However, deploying these models on resource-constrained hardware presents challenges in terms of performance, stability, and memory usage, which remain underexplored in real-world scenarios. This study proposes a replicable evaluation pipeline to identify SLMs suitable for deployment on resource-constrained Android edge devices, using mainly Android TV as a low-resource platform. The first phase evaluates SLMs (65M–1B parameters) on a smart TV with an ARM Cortex-A55 processor, collecting performance and quality metrics. This phase reveals critical constraints related to memory footprint and inference quality. The second phase presents a comparative analysis of inference acceleration techniques across consumer Android devices, such as TVs and smartphones. It systematically evaluates the impact of low-level linear algebra libraries and various inference engine implementations executed on Android TV. Results show that choosing the right inference engine—especially MNN, in our context—yields up to 12× speedups over alternatives. These findings offer actionable guidance for deploying SLMs in production on constrained devices. The pipeline supports automation and reproducibility, aiding real-world integration of generative AI at the edge.