High-fidelity simulation software underpins design optimization of complex engineering systems such as nuclear reactors and aerospace vehicles, yet these tools remain difficult to access and extend, and each evaluation can be computationally expensive, which collectively impedes rapid, iterative design. In current practice, domain experts must manually formulate optimization problems, choose algorithms, and tune hyperparameters and constraints, leading to brittle, low-throughput trial-and-error workflows that limit exploration in high-dimensional design spaces. We present AICO (AI Core Optimus), a large language model (LLM)-driven framework for fully automated optimization around "black-box" industrial solvers: AICO interacts with engineers through natural language, translates design intent into formal optimization tasks, and autonomously selects, configures, and schedules algorithms from multiple optimization libraries without modifying the underlying simulation codes. Across representative industrial benchmarks—including heat conduction, fluid machinery component design, and nuclear reactor core eigenvalue calculations—AICO markedly reduces human intervention and wall-clock time while maintaining or improving solution quality relative to expert-crafted baselines, and it robustly addresses nonlinear, high-dimensional problems where fixed single-algorithm pipelines often stagnate. By coupling high-fidelity simulations with LLM-based meta-level optimization, AICO establishes a general "AI + simulation optimization" paradigm that enables intelligent automation to be embedded into existing engineering workflows, improving Research and Development efficiency and accelerating industrial innovation.