This work presents a fully thermodynamically consistent, data-driven constitutive framework in which plastic admissibility, flow direction, and hardening evolution are governed by a neural-network-based representation of the recoverable free energy. The model is formulated directly at the level of the Helmholtz potential and enforces the dissipation inequality through a state-dependent energy ceiling $\psi^{\max}$ that defines the elastic domain. Unlike classical yield functions, $\psi^{\max}$ is learned from data and depends on a set of internal variables allowing to capture isotropic hardening, Bauschinger effect, and Lode-angle dependency. This allows the model to reproduce a wide spectrum of elastic-plastic behavior with a single manifold. The constitutive update is obtained by solving the incremental maximum-dissipation problem subject to the admissibility constraint $\psi \le \psi^{\max}$. A two-residual Newton return-mapping scheme is developed, in which the first residual enforces stress consistency and the second enforces the active energy ceiling condition. Automatic differentiation of the neural network provides the required gradients with respect to stress and internal variables, enabling a fully consistent Jacobian for robust and quadratic convergence. Bauschinger effect is incorporated through additional stored energy terms governing the recoverable part of the plastic work in stress space, allowing the model to accurately capture non-proportional loading, and path dependence. The resulting formulation unifies elastic energy storage, yield surface evolution, flow direction, and hardening mechanisms within a single energy-based structure, while guaranteeing non-negative dissipation for all admissible processes. Extensive multiaxial benchmark tests, covering monotonic, cyclic, and non-proportional loading paths, demonstrate that the model achieves high predictive accuracy, numerical stability, and full thermodynamic admissibility. The proposed framework offers a generalizable alternative to hand-crafted plasticity models, providing a flexible yet physically constrained approach for integrating experimental or simulation data into constitutive modeling.