CUDA for Big Data + machine learning

Hello, I have developed a state of the art engine for machine learning and I want to improve it. Some of its sections are written in CUDA and can be improved. I have some ideas on how to improve the code but I need some one to test them and potentially come up with new ideas on his/her own. The current challenge and bottleneck involves performing sub-sums of a very large array of doubles where the sub-sums to be performed are described in a series of index arrays. The person needs to have good knowledge of CUDA and parallel algorithms and preferably be proficient in libraries like Thrust. This project will involve several iterations and potentially several sykype/phone calls to discuss its progress and possible directions. The amount of code to be written is probably going to be very small in the end but the number of ideas to try to find the optimal code might large.