Building AI Systems That Work Offline: An Offline-First Architecture and Feasibility Assessment for Low-Connectivity Environments

Abstract

Artificial intelligence applications increasingly rely on continuous cloud connectivity for inference, data synchronization, and model updates. This dependency limits their applicability in regions where Internet access is intermittent, costly, or unavailable, such as rural schools and other resource-constrained settings. This work proposes an offline-first reference architecture for intelligent applications that combines local inference, persistent on-device storage, connectivity monitoring, deferred synchronization, a local explanation layer, and optional cloud services, so that essential functionality is preserved during connectivity disruptions. Ten design requirements are derived for this class of systems, and a two-level evaluation framework is defined that separates model-level metrics from system-level resilience metrics. As a first feasibility assessment, three local language-model configurations were benchmarked on a commodity desktop computer without a dedicated GPU and with all network access disabled: Llama 3.2 3B Instruct in FP16, the same model quantized to 4 bits (Q4_K_M), and Gemma 2B. On a 50-item multiple-choice test in Spanish, the quantized model reduced model size by 68.6%, peak memory use by 49.2%, and mean latency by 36.5% relative to the FP16 version, while achieving comparable accuracy (82% versus 86%). Gemma 2B required the fewest resources but reached only 60% accuracy and showed a marked bias toward the first answer option. These results indicate that local inference with small quantized language models is technically feasible on widely available hardware, supporting the viability of the proposed architecture. The system-level evaluation of a full prototype under emulated connectivity conditions is defined as future work.