Accurate aerodynamic prediction remains challenging due to the prohibitive computational cost of High-Fidelity (HF) simulations, which severely limits their direct application in data-driven modelling. This work presents a multi-fidelity deep learning framework that integrates unsupervised representation learning via autoencoders with transfer learning methodologies to bridge the gap between information sources. This core architecture is coupled with a newly developed Multi-Split Conformal Prediction (MSCP) strategy to achieve uncertainty-aware aerodynamic data fusion under extreme data scarcity. The proposed model leverages abundant Low-Fidelity (LF) data to pre-train an encoder, extracting a compact latent physics representation that serves as a frozen knowledge base. The decoder is subsequently fine-tuned using scarce HF samples, while a dedicated adaptation layer corrects LF-to-HF biases and bridges dimensionality differences. To ensure reliability, the MSCP strategy extends the conformal prediction framework to provide uncertainty quantification. By employing a multi-split resampling scheme, the methodology allows the final model to be trained on the entire HF dataset, thus preventing the data-efficiency penalty of a fixed calibration split. The framework was validated on surface-pressure distributions for NACA airfoils and an XRF1 transonic wing. Results demonstrate high predictive accuracy with minimal HF data; for the wing case, the model achieved an R2 = 0.965 using only a fraction of the potential HF database. Furthermore, the MSCP framework yields robust uncertainty bands with pointwise coverage exceeding 95%. These bands adaptively widen in regions of complex nonlinearities, such as shock waves, providing actionable confidence margins for engineering decisions. By combining extreme data efficiency with uncertainty quantification, this work offers a scalable and reliable solution to accelerate aerodynamic design cycles in data-scarce environments.