Inverse rendering is a challenging problem due to the inherent ambiguity in recovering 3D geometry, materials, and lighting from multi-view RGB images. Recent methods represent geometry and materials as neural networks, recovering them through an analysis-by-synthesis approach. However, accurately modeling shadows and inter-reflections remains a significant obstacle. In this paper, we propose a two-stage inverse rendering pipeline that effectively reconstructs geometry, materials, and illumination. First, we decompose the neural radiance field into diffuse and specular components and introduce an occlusion network to jointly improve surface reconstruction quality. We then use Monte Carlo-based path tracing to model shadows and inter-reflections. To enhance efficiency and quality, we employ separate hash-encoded MLPs for geometry and material representation. Experiments demonstrate that our method outperforms previous work on synthetic and real datasets, accurately recovering geometry, roughness, and high-quality albedo, while supporting realistic re-rendering and relighting.