Signals from sensors are often analyzed in a sequence of steps, starting with the raw sensor data and eventually ending up with a classification or abstraction of these data. This paper will give a practical example of how the same information can be trained and used to initiate multiple interpretations of the same data on different, application-oriented levels. Crucially, the focus is on expanding embedded analysis software, rather than adding more powerful, but possibly resource-hungry, sensors. Our illustration of this approach involves a tangible input device the shape of a cube that relies exclusively on lowcost accelerometers. The cube supports calibration with user supervision, it can tell which of its sides is on top, give an estimate of its orientation relative to the user, and recognize basic gestures.