Sensor networks have become an important source of data with numerous applications in monitoring various real-life phenomena as well as industrial applications and traffic control. Unfortunately, sensor data is subject to several sources of errors such as noise from external sources, hardware noise, inaccuracies and imprecision, and various environmental effects. Such errors may seriously impact the answer to any query posed to the sensors. In particular, they may yield imprecise or even incorrect and misleading answers which can be very significant if they result in immediate critical decisions or activation of actuators. In this paper, we present a framework for cleaning and querying noisy sensors. Specifically, we present a Bayesian approach for reducing the uncertainty associated with the data, that arise due to random noise, in an on-line fashion. Our approach combines prior knowledge of the true sensor reading, the noise characteristics of this sensor, and the observed noisy reading in order to obtain a more accurate estimate of the reading. This cleaning step can be performed either at the sensor level or at the base-station. Based on our proposed uncertainty models and using a statistical approach, we introduce several algorithms for answering traditional database queries over uncertain sensor readings. Finally, we present a preliminary evaluation of our proposed approach using synthetic data and highlight some exciting research directions in this area.