Modern software-defined networks, such as Open Radio Access Network (O-RAN) systems, rely on artificial intelligence (AI)-powered applications running on controllers interfaced with the radio access network. To ensure that these AI applications operate reliably at runtime, they must be properly calibrated before deployment. A promising and theoretically grounded approach to calibration is conformal prediction (CP), which enhances any AI model by transforming it into a provably reliable set predictor that provides error bars for estimates and decisions. CP requires calibration data that matches the distribution of the environment encountered during runtime. However, in practical scenarios, network controllers often have access only to data collected under different contexts -- such as varying traffic patterns and network conditions -- leading to a mismatch between the calibration and runtime distributions. This paper introduces a novel methodology to address this calibration-test distribution shift. The approach leverages meta-learning to develop a zero-shot estimator of distribution shifts, relying solely on contextual information. The proposed method, called meta-learned context-dependent weighted conformal prediction (ML-WCP), enables effective calibration of AI applications without requiring data from the current context. Additionally, it can incorporate data from multiple contexts to further enhance calibration reliability.