We present a novel approach to constrained data assimilation that utilizes the ensemble transform Kalman filter (ETKF) for improved state estimation in complex dynamical systems. Our method enforces hard constraints directly within the optimization process using quadratic programming. This leads to physically consistent solutions and enhances the robustness of the data assimilation process. After validating the constrained ETKF, we generate ensembles of state variables, which serve as training data for a reinforcement learning (RL) framework. The RL model is designed to estimate the ensemble states both accurately and efficiently while maintaining adherence to physical constraints. By integrating data assimilation with RL, we aim to enhance the physical fidelity of state estimation in systems governed by complex dynamics and strict constraints. This hybrid approach offers significant potential for advancing predictive accuracy and computational efficiency in uncertain environments. To show the method's performance, we provide examples such as the Burgers' equation.