In many peer-to-peer sensor network architectures, the signal measurements acquired at nodes are quantized, and the associated bit representations are propagated through the network to data fusion nodes where inference is performed. From this perspective, the quantizer design problem can be viewed as one of coding for estimation with communication constraints. A naive approach would have each node design its quantization without regard for the network structure. However, a substantially more resource efficient network results from the quantizer taking into account that the intermediate nodes propagating these bit representations are receiving their own measurements as well as bit representations from other nodes. From an information-theoretic perspective, this these intermediate nodes have so-called side information that can be exploited in the code design. We develop powerful greedy coding algorithms based on this principle, and show how they achieve or approach fundamental limits in prototype topologies. From an architectural perspective, the new algorithm represents the design of a network-aware application layer for this problem, which our results establish is substantially more resource-efficient than a traditional layered architecture.