The latest interesting open-source project we have come across is Likely, an image recognition run-time for heterogeneous architectures. Likely provides a domain-specific language for dealing with image recognition and is optimized to run on both CPUs and GPUs.
Likely is designed around heterogeneous architectural support, an efficient syntax for feature representation and statistical learning, immediate visual feedback during algorithm development, can be embedded into other projects/languages, and is free/open-source software under the Apache 2.0 license.
The Likely Kernel Library is built atop LLVM and uses its MCJIT just-in-time compiler while supporting multi-core CPUs via custom OpenMP back-end and also supporting GPUs and other accelerators via OpenCL, HSAIL, and CUDA. LLVM is responsible for all optimization tasks for this image recognition DSL.
Should you be interested in image recognition or just interesting LLVM projects, checkout LibLikely.org