This study investigates the feasibility of reconstructing electron temperature and density profiles on the COMPASS tokamak using neural networks trained on soft x-ray and absolute extreme ultraviolet radiation diagnostics, complemented by selected global and local plasma measurements. The proposed approach enables kinetic profile inference with microsecond temporal resolution, far exceeding that of the standard Thomson scattering diagnostic. The model performance is evaluated for different diagnostic combinations and plasma regimes, including L-mode, H-mode. Typical reconstruction errors remain within about 10% in the plasma core and increase to approximately 20% near the edge. The method is further shown to capture temperature and density evolution during fast transient phenomena such as sawtooth crashes and edge-localized modes, demonstrating its capability to follow rapid events and serve as a potential tool for their detailed analysis.