The goal is improving performance and efficiency of ML (facial-recognition) workload on resource-limited environment by offloading resource-intensive work to the cloud and moving as many components as possible to the kernel space.
The diagram above roughly covers the interactions between different modules. However, it didn't reveal the whole picture. There are still some nitty-gritty details under the hood.