Accurate estimation of motor winding temperature is essential for reliable operation and thermal protection of electric machines. Direct temperature measurement, however, is often impractical due to sensor placement limitations and harsh operating conditions. Consequently, data-driven temperature estimation methods have emerged as an effective alternative to conventional thermal modeling. This study presents a compact data-driven framework for motor winding temperature estimation with an emphasis on deployment in embedded systems. A supervised learning model is developed to estimate winding temperature from operational signals, capturing nonlinear thermal behavior under varying conditions. The model operates without direct temperature measurements while maintaining robustness to noisy inputs. The trained model is then systematically reduced to satisfy strict memory and computational constraints required for microcontroller implementation. A general model reduction strategy is applied to minimize model size while preserving estimation accuracy. Lightweight model configurations are evaluated to identify designs that meet tight memory limits and support stable real-time inference. The results demonstrate that accurate motor winding temperature estimation can be achieved using compact models suitable for embedded deployment. The proposed approach enables efficient thermal monitoring in electric machine applications with limited computational resources.