Deep reinforcement learning (DRL) shows great potential in irrigation scheduling. However, deploying complex DRL models on field hardware remains difficult. This paper proposes a deployable smart irrigation system. We integrate a biophysical crop growth model with an agent. A modular Python-Flask backend enables efficient real-time inference. We decoupled inference from training to reduce computational overhead. The system uses a bidirectional long short-term memory encoder to extract temporal features. We developed a browser/server (B/S) architecture dashboard for real-time monitoring and explainable decision heatmaps. System-level validation was conducted using historical data from Xinjiang cotton fields. The results show sub-10 ms inference latency and a 0.45 s total system response time. The system provides a 1.9% yield and 5.9% irrigation water use efficiency (IWUE) boost over standard soft actor-critic (SAC) algorithms, proving its algorithmic efficacy. This architecture transforms theoretical gains into executable irrigation schedules for real-world farming. It ensures stable and failure-free operation across continuous growing seasons.