SUSTAINABLESIGHT: a resource-aware vision–edge AI architecture for adaptive navigation and STEM accessibility

Assistive artificial intelligence systems for visually impaired users commonly address mobility and educational accessibility as separate computational tasks, resulting in fragmented architecture, redundant inference, and inefficient use of computational resources. This study presents SUSTAINABLESIGHT, a resource-aware vision–edge AI architecture intended for wearable deployment that integrates navigation perception and mathematical-expression recognition through shared, context-aware execution. The proposed architecture supports navigation (NAV), learning (LEARN), and hybrid (HYBRID) operating modes and employs a mode-aware scheduler to activate only the computational components required by the current task. A reproducible computational evaluation was conducted using public datasets and measured inference traces on a GPU runtime; no physical wearable device was evaluated. For navigation perception, YOLOv8n evaluated on the COCO 2017 validation set achieved a precision of 0.6343, recall of 0.4746, mAP@50 of 0.5188, and mAP@50–95 of 0.3681, with a mean inference latency of 7.24 ms on an NVIDIA Tesla T4 GPU. Mathematical-expression recognition using pix2tex was evaluated on a fixed 500-expression subset of an Im2LaTeX validation dataset; 499 expressions were evaluable, with a mean normalized similarity of 0.5940, median similarity of 0.6394, and median character error rate of 0.4343. The substantially greater computational cost of mathematical recognition motivated adaptive branch activation. In a 1,000-request trace-driven workload, SUSTAINABLESIGHT maintained 100% request coverage while reducing model invocations from 2,000 to 1,098 (45.10%) and mean request latency from 3562.22 ms to 1311.32 ms (63.19%) relative to a fixed unified baseline. Mean-latency reductions of 16.15–81.61% were observed across five workload compositions, and scheduler ablation showed that avoiding unnecessary mathematical-recognition inference accounts for most of the latency improvement. Sustainability is assessed here as computational resource efficiency; device energy consumption and carbon emissions were not measured. These results indicate that SUSTAINABLESIGHT offers a reproducible, resource-aware basis for integrating mobility and STEM-accessibility workloads, while improved mathematical recognition, deployment of wearable edge hardware, and user validation remain important directions for future work.