The violation of a Bell inequality is the paradigmatic example of device-independent quantum information: The nonclassicality of the data is certified without the knowledge of the functioning of devices. In practice, however, all Bell experiments rely on the precise understanding of the underlying physical mechanisms. Given that, it is natural to ask: Can one witness nonclassical behavior in a truly black-box scenario? Here, we propose and implement, computationally and experimentally, a solution to this ab initio task. It exploits a robust automated optimization approach based on the stochastic Nelder-Mead algorithm. Treating preparation and measurement devices as black boxes, and relying on the observed statistics only, our adaptive protocol approaches the optimal Bell inequality violation after a limited number of iterations for a variety photonic states, measurement responses, and Bell scenarios. In particular, we exploit it for randomness certification from unknown states and measurements. Our results demonstrate the power of automated algorithms, opening a venue for the experimental implementation of device-independent quantum technologies.