I assume that by "proper hypervisor", you mean a hypervisor that runs a full kernel. This project does use a hypervisor, but apparently only to set up memory mapping and trap the syscalls for translation. Hypervisors that run a full kernel typically end up emulating hardware host interfaces, which leads to its own set of challenges and often introduces overhead that isn't present when emulating at the kernel ABI level (e.g. operations going through both the guest and host VFS/FS/SCSI stacks). I don't think it's obvious that one approach is clearly superior to the other overall.
