Additive manufacturing (AM) enables the fabrication of lightweight lattice structures with exceptional stiffness-to-weight ratios; however, optimizing functionally graded cellular designs for specific load cases remains computationally prohibitive due to the reliance of traditional topology optimization on costly iterative simulations [1]. Recent studies have adopted AI-based surrogate models to accelerate this process; yet existing frameworks typically optimize only simple geometric parameters (e.g., unit cell size and strut thickness) and rely on large offline datasets generated by commercial finite element software for model training [2,3]. To address these limitations, this work introduces a standalone lattice optimization framework that integrates a native Finite Element Analysis (FEA) engine with a stress-informed, multi-objective Bayesian optimization strategy to identify optimal density-graded lattice infill structures. The framework is driven by FEA results and extracted von Mises stress scalar fields, which are used for density grading of the lattice architecture and for selecting the optimal unit cell size and lattice topology. Unlike methods that rely on datasets for surrogate model training, this framework employs an active learning approach in which an AI surrogate model dynamically approximates the structural response surface in real time. The optimization is formulated using a weighted multi-objective function designed to maximize the lattice stiffness-to-weight ratio while simultaneously enforcing strict mechanical failure constraints. The Bayesian algorithm intelligently navigates the design space through an acquisition function that balances exploration of topological parameters with exploitation of known high-performance regions. Candidate structures are iteratively validated by the native FEA solver, which provides mechanical feedback to refine the surrogate model. This synergistic approach enables rapid convergence toward optimal, functionally graded lattice infills, ensuring mechanical integrity and topological efficiency without the computational burden or external dependencies of traditional simulation software.