This paper explores a future in which drones serve as extensions to cellular networks. Equipped with a WiFi interface and a (LTE/5G) backhaul link, we envision a drone to fly in and create a WiFi network in a desired region. Analogous to fire engines, these drones can offer on-demand network service, alleviating unpredictable problems such as sudden traffic hotspots, poor coverage, and natural disasters. While realizing such a vision would need various pieces to come together, we focus on the problem of "drone placement". We ask: when several scattered users demand cellular connectivity in a particular area, where should the drone hover so that the aggregate demands are optimally satisfied? This is essentially a search problem, i.e., the drone needs to determine a 3D location from which its SNR to all the clients is maximized. Given the unknown environmental conditions (such as multipath, wireless shadows, foliage, and absorption), it is not trivial to predict the best hovering location. We explore the possibility of using RF ray tracing as a hint to narrow down the scope of search. Our key idea is to use 3D models from Google Earth to roughly model the terrain of the region, and then simulate how signals would scatter from the drone to various clients. While such simulations offer coarse-grained results, we find that they can still be valuable in broadly guiding the drone in the right direction. Once the drone has narrowed down the 3D search space, it can then physically move to quickly select the best hovering location. Measurement results from a WiFi mounted drone, communicating with 7 clients scattered in the UIUC campus, are encouraging. Our early prototype, DroneNet, reports 44% throughput gain with only 10% measurement overhead compared to a full scan of the entire region.