Democratiassess: Edge Native Quantized Ai for Equitable, Offline Adaptive Exams in Resource-constrained Universities

The persistent digital divide excludes students in resource-constrained regions from Computerized Adaptive Testing (CAT), which requires stable internet for real-time, serverbased inference. This paper introduces DemocratiAssess , an edge-native framework that enables high-fidelity, offline adaptive assessment. By combining a novel offline-first architecture with extreme 4-bit integer quantization (INT4) with SIMD-optimized bitunpacking kernels and knowledge distillation, we compress transformer models for local deployment on affordable Raspberry Pi hardware, eliminating cloud dependency. Experimental results demonstrate that DemocratiAssess achieves an average system latency of 38.2 ms (σ = 4.1ms, P95=52ms) and a memory footprint of 150 MB, representing a 97% reduction compared to cloud-based systems, while retaining 96.2% of the baseline model’s psychometric accuracy and reducing measurement disparity by 85.7% through Psychometric-Aware Quantization (PAQ). Three-year total cost of ownership (TCO) is reduced to $90.12 including hardware and electricity. This work establishes the technical and economic feasibility of equitable, personalized assessment in low-connectivity academic environments, offering a sustainable model for educational technology in the Global South.