Embedded systems can already capture data produced at high rates, and embedded CPU and sensor performance are still rapidly improving. Radio technology, however, cannot keep pace, and will not in the future due to known physical limits of shared communication channels. This leads to a fundamental gap between the data a sensor network node can collect and the data it can transmit back for analysis. VanGo, our software system for data collection, uses flexible transcoding to narrow this gap. To make effective use of channel bandwidth, data reduction software must run on sensor nodes. However, to calibrate how data reduction software should run, that same software should be capable of running on the back end on real data received from the network. In VanGo, users decide where data processing occurs. To show that transcoding helps, we evaluate two radically different applications: acoustic collection and the measurement of neural activity. Among our findings is that in bandwidth-limited environments, proactive filtering of some of our signal can result in collecting three times the signal energy than we could by removing silent periods alone.