Construction remains among the most hazardous sectors in India, where workers bear a share of global occupational hazards far exceeding their share of the labour force. Automated visual monitoring offers continuous supervision, but existing systems rely on convolutional detectors demanding large annotated datasets and specialist expertise, placing them beyond the reach of small and medium contractors. Recent multimodal language models have been evaluated only as benchmarks on static images, without a deployable architecture or any means of calibrating alerts to site conditions. This paper presents a four-layer framework integrating commodity Wi-Fi cameras with a compact hybrid convolutional-vision transformer backbone, a rule-mapping module that encodes site safety requirements as visually verifiable checks, and a preference-alignment loop in which routine confirm-or-dismiss responses from the site engineer are converted into preference pairs and used to calibrate alert severity through direct preference optimization. Three algorithms specify adaptive frame sampling, hazard severity ranking and onsite alignment. Compared with manual supervision, wearable sensing, convolutional detectors and cloud inference, the design requires no site-specific labelled dataset from the adopting contractor and is reconfigured by editing a table rather than by retraining. Limitations and field validation requirements are stated.

Construction safety, Vision transformer, Preference alignment, Site monitoring, Personal protective equipment.