Wildfire risk assessments across large geographical regions often rely on simulation-based fire models, with few approaches that leverage physical event characteristics. As climate change intensifies wildfire activity, there is an increasing need for scalable risk assessment frameworks grounded in empirical observations. This study presents a novel, national-scale wildfire risk model using the VIIRS 375 m active fire satellite product. While the VIIRS product provides high-resolution thermal anomaly detections, limitations within its current flagging algorithms and false-positive sorting has limited its use for risk analysis. To address this, a systematic filtering algorithm was developed to independently assess ten years of VIIRS detections across various criteria associated with false-positive thermal anomalies, resulting in a wildfire-focused dataset. A transitive, spatio-temporal clustering approach was implemented to reconstruct events out of individual detections. Events were evaluated for wildfire probability using physical characteristics including burned area, expansion rate, duration, and false detection proportion. High- and low-confidence events were used to train a machine learning ensemble in which a national wildfire dataset was constructed and predicted using a separate set of event characteristics. Event-level intensity scores were aggregated into a composite, gridded wildfire risk model of the continental U.S. at three different resolutions and model divergence was assessed among existing risk products. This framework provides a scalable, empirical alternative to traditional wildfire risk models, and demonstrates the efficacy of integrating satellite-based observation into large-scale wildfire risk assessment.