This paper introduces a software and hardware-integrated method for handwritten digit recognition using a customized Convolutional Neural Network (CNN architecture. The process comprises two concurrent tracks hardware implementation and software model development. The CNN was first designed and trained with Python frameworks before being transformed into a hardware-compatible format. It was then put on a Raspberry Pi (RPB) Board and converted into an ASIC design using Cadence tools. The conversion of floatingpoint to fixed-point representation of neural network parameters, such as weights and biases, allows for a smooth integration into a single hardware module. The RPB is used to verify the model’s performance, and the Vivado platform is used for further verification, it uses comprehensive test cases to guarantee correctness and resilience. Moreover, a comparative study of the execution timings of hardware implementations (RPB) and software platforms (CPU, GPU) is carried out, emphasizing the design efficiency and interface performance. This method shows that it is feasible to implement neural networks with optimal performance on hardware with limited resources.