The exponential proliferation of digital misinformation presents a critical challenge for information integrity, particularly in developing regions where network latency, data costs, and intermittent connectivity prohibit reliance on cloud-based verification systems. While Large Language Models (LLMs) and Transformer-based architectures (e.g., BERT) offer state-of-the-art performance in automated fact-checking, their significant memory footprint (>400 MB) and computational latency render them unsuitable for client-side deployment on consumer hardware. This study addresses this "Digital Divide" by proposing a lightweight, offline-capable architecture for real-time misinformation detection. Leveraging the LIAR dataset, I fine-tuned a DistilBERT model and engineered a compression pipeline utilizing Dynamic Quantization (INT8) and ONNX (Open Neural Network Exchange) Runtime Optimization. The approach achieved a 74.8% reduction in model size (from 255.45 MB to 64.45 MB), successfully crossing the critical 100 MB threshold required for browser extension deployment. Furthermore, inference latency on standard CPU hardware was reduced by 55.2% (from 52.73 ms to 23.58 ms), establishing feasibility for synchronous user interaction. These results demonstrate that complex Natural Language Processing (NLP) tasks can be democratized for edge deployment, enabling privacy-preserving, accessible AI safety tools in regions with limited connectivity.