In the Sci-Fi-Turbo project, a multi-fidelity optimization framework is developed to systematically integrate scale-resolving simulations (SRS) into the turbomachinery aerodynamic design process. The framework combines Reynolds-averaged Navier–Stokes (RANS) and SRS-based flow predictions to balance computational cost and predictive accuracy and is demonstrated for the automated optimization of a compressor cascade. The optimization objective is the reduction of profile losses, while constraining the outflow angle at the aerodynamic design point (ADP) and at two off-design incidences of ±5°. To achieve this objective, a two-fidelity optimization strategy is employed. The low-fidelity (LF) model is based on RANS simulations using the Menter SST k–ω turbulence model in combination with the γ–Reθ transition model, whereas high-fidelity (HF) data are obtained from large-eddy simulations (LES) using a high-order discontinuous Galerkin (DG) solver. The numerical setup exploits pitchwise and spanwise periodicity in a quasi-three-dimensional configuration. The inflow is characterized by a Reynolds number of Re=150 000, a Mach number of M=0.6, and synthetically generated isotropic turbulence in the LES. The optimization involves 15 geometric design variables and requires the evaluation of approximately 300 candidate geometries to achieve a converged Pareto front. To ensure tractability at SRS level, the workflow emphasizes a high degree of automation, including automatic transition detection, robust convergence monitoring, and uncertainty-aware evaluation of LES results to control the noise in the surrogate model. A comparative analysis shows that RANS-based optimization can yield Pareto-optimal designs that exhibit significant performance degradation under off-design conditions and may violate kinematic constraints, when reassessed using LES~\cite{goinis2026}. These findings motivate the Sci-Fi-Turbo strategy of multi-fidelity optimization, in which surrogate models trained on extensive LF data are selectively enriched with sparse HF samples. The presented framework establishes the methodological and computational foundation required for subsequent data-driven and adaptive extensions of SRS-informed turbomachinery optimization.