Designing optimal control pulses that drive a noisy qubit to a target state is a challenging and crucial task for quantum engineering. In a situation where the properties of the quantum noise affecting the system are dynamic, a periodic characterization procedure is essential to ensure the models are updated. As a result, the operation of the qubit is disrupted frequently. In this paper, we propose a protocol that addresses this challenge by making use of a spectator qubit to monitor the noise in real-time. We develop a quantum machine-learning-based quantum feature engineering approach for designing the protocol. The complexity of the protocol is front-loaded in a characterization phase, which allow real-time execution during the quantum computations. We present the results of numerical simulations that showcase the favorable performance of the protocol.