I'm working in industry and basically use computational science as part of my daily work. I have done work on the "numeric_bindings" library, but mainly from the maintenance, portability, support and testing point of view. (The most recent development work was done by Rutger ter Borg.)
The real world application of computational science is great, especially when you work with actual measurement data. The models are often oversimplified, but still manage to reproduce and predict real world behavior surprisingly accurate. In a certain sense, the whole thing is unbelievably messy, if you compare it to the idealized academical world. Many problems that academia proves to be unsolvable (in an ideal sense) can still be solved accurate enough for practical purposes. For me, the idealized academical world is always a nice repose from this real world, and I think it is really important as a counterweight.