This work presents HOMEY, a practical and scalable computer vision framework designed to automate property risk assessment in insurance and real estate contexts. By augmenting YOLO with heuristic object masking, the model amplifies subtle and weak signals often obscured in cluttered real-world scenes. In addition, a custom risk-aware loss function is introduced to handle class imbalance while incorporating severity-aware weighting, enabling more meaningful predictions for downstream decision-making. Evaluated across 17 diverse risk categories—including structural issues, maintenance deficiencies, and safety hazards—HOMEY demonstrates strong performance gains in both detection accuracy and reliability compared to conventional YOLO baselines. Beyond raw performance, the framework emphasizes interpretability and operational efficiency, making it suitable for deployment in large-scale underwriting pipelines. This study highlights the growing role of domain-aware architectural enhancements in bridging the gap between academic computer vision models and real-world insurance applications.
