The AI-Native Air Interface model aims to revolutionize the modern approach to communication systems by introducing, for the first time, a fully data-driven deep learning approach to the design of the physical layer. Most existing methods for modulation and decoding of RF signals rely on handcrafted algorithms that often perform poorly in the presence of noise, fading, and dynamic channel variations. In this work, the authors introduce a novel AI-Native architecture, based on the ConvMixer and MLP-Mixer frameworks, that can learn end-to-end modulation and coding schemes. This approach combines an encoder, differentiable channel simulation, and decoder pipeline, implemented in Python using TensorFlow and PyTorch. This encoder efficiently extracts robust signal features, adds realistic channel noise, and reconstructs or classifies the signal with high precision at the decoder. Experimental results have shown a modulation accuracy of 94.2%, an average reconstruction MSE of 0.023, and latent robustness of 0.95, outperforming traditional systems. This research will help next-generation wireless networks, IoT communication, and AI-driven radio systems achieve adaptive, resilient signal transmission in uncertain environments. The proposed model serves as the basis for intelligent, self-optimizing communication frameworks that can evolve in response to changing conditions and spectrum demands.