This diploma thesis presents the design, implementation, and evaluation of a system for real-time neural emulation of nonlinear audio effects. The proposed HybridAmpModel architecture follows a hybrid structured black-box approach with an L-NL-L topology physically motivated by the signal chain of a guitar amplifier. It consists of a trainable front-end FIR filter, a nonlinear core based on a Temporal Convolutional Network (TCN) with exponentially dilated causal convolutions, and FiLM (Feature-wise Linear Modulation) conditioning on the gain parameter using the Snake activation function. The model is trained on real recordings of two target devices — the digital plugin SGA1566 and the analog tube guitar amplifier Laney Ironheart — using a combined loss function comprising ESR (Error-to-Signal Ratio) and MR-STFT (Multi-Resolution Short-Time Fourier Transform) with a pre-emphasis filter. Four depth variants of the architecture were trained (5, 7, 8, and 10 layers). The trained models are exported using the RTNeural library and deployed as functional VST3 plugins built on the JUCE 8 framework with zero algorithmic latency. The best objective results were achieved by the SGA1566 10L variant (validation ESR 8.7\%) and the Laney Ironheart 8L variant (ESR 53.8\%). A subjective online listening test (29 respondents, MUSHRA, ABX, and paired preference methods) achieved an average MUSHRA score of 55.4 points. The preference test results showed that the overall preference for the original recording (56.2\%) was not statistically distinguishable from random selection. Moreover, for the SGA1566 8L and Laney 5L variants, respondents actually preferred the model's prediction over the original. The generated sound, despite measurable deviations, possesses high aesthetic value and is perceived by listeners as musical and practically usable.