We consider the problem of localizing a wireless client in an indoor environment based on the signal strength of its transmitted packets as received on stationary sniffers or access points. Current state-of-the art indoor localization techniques have the drawback that they rely extensively on a 'training phase'. This 'training' is a labor intensive process and must be done for each target-area under consideration for various device types. This clearly does not scale for large target areas. The introduction of unmodeled hardware with heterogeneous power-levels etc further reduces the accuracy of these techniques. We propose a solution in which we model the received signal strength as a Gaussian Mixture Model (GMM). We use expectation maximization to find the parameters of our GMM. We can now give a location fix for a transmitting device based on the maximum likelihood estimate. This way, we not only avoid the costly 'training phase' but also make our location estimates much more robust in the face of various form of heterogeneity and time varying phenomena. We present our results on two different indoor testbeds (CEWIT and Computer Science Buildings in Stony Brook University) with multiple WiFi devices (iphones, android phones, laptops, netbooks). We demonstrate that the accuracy is at par with state-of-the-art techniques but without requiring any training. We also show an application of such localization in extracting the hidden social structure of the occupants of the building based on their WiFi activity. We show interesting observations from the Computer Science building in Stony Brook University.