In this work, we propose a novel method for low-light image enhancement directly operating on RAW sensor data towards modeling a neural Image Signal Processor (ISP). Low-light photography presents significant challenges due to the presence of noise and limited visibility, which degrade image quality and obscure fine details. Traditional enhancement methods often struggle to balance noise suppression with the preservation of structural and color fidelity. Specifically, we process Canon CR2 RAW images using a modular pipeline composed of Decom-Net for image intrinsics decomposition and the Hierarchical Noise-Deinterlace Network (HNN) for reflectance denoising, illumination enhancement, and final image intrinsics fusion. Our approach begins by decomposing the low-light input into reflectance and illumination components (referred to as image intrinsics in this work) using Decom-Net, enabling independent and targeted processing of noise and exposure artifacts. The HNN is subsequently employed in three stages: to denoise the reflectance, enhance the illumination, and fuse the refined intrinsics components into a high-quality output. This decomposition-guided, multi-stage framework facilitates adaptive enhancement across diverse low-light scenarios. Experimental results demonstrate the proposed method significantly outperforms conventional approaches, yielding images with enhanced brightness, reduced noise, and preserved natural color and detail.