Automotive radar is essential for enabling advanced driver assistance and autonomous driving. Yet, precise direction-of-arrival (DoA) estimation and robust interference mitigation remain challenging on embedded radar System-on-Chip (SoC) platforms due to limited compute and memory. This work presents a streamlined signal processing pipeline that combines classical radar techniques with lightweight AI models to enhance DoA accuracy and interference mitigation on a 4T4R MIMO radar SoC. By fusing analytical methods with efficient deep learning architectures, the solution achieves high angular resolution within the constraints of embedded DSPs. The implementation builds on available NXP radar chips, demonstrating how advanced radar perception can be realized on cost-sensitive automotive hardware.