How would you optimize a data processing pipeline that is currently running out of memory at AI Camp?