This study presents a data-driven framework for estimating transonic aerodynamic loads using a multi-fidelity approach built on Bayesian neural networks (BNNs) and transfer learning (TL). The proposed model, MF-BayNet, integrates low-, mid-, and high-fidelity aerodynamic data to achieve accurate predictions with robust uncertainty quantification while significantly reducing the need for expensive high-fidelity simulations. The methodology is tested on two challenging cases: a transonic wing and a complex eVTOL configuration. Results show that MF-BayNet outperforms classical surrogate modeling techniques such as Co-Kriging in both accuracy and generalization, especially under strong nonlinearities. Through BNNs, the model inherently captures both epistemic and aleatoric uncertainty, providing reliable confidence estimates