I had played with this idea some time back and gave up after some very specific flaws came became clear.

The good probability that a 5 digit combination is found in Pi will be in the range of locations above 10000, for example I once located by 6 digit phone number in position 685214 which was not actually helpful at all.

Further we are not sure if Pi is normal hence the better idea would be use a simple computable normal series.

It was just yesterday I uploaded a paper that presents a idea for Compressing Random Data to -> https://www.academia.edu/7620004/Advanced_Compression_Techni... which proposes an Idea to push multiple bytes represented by a positions in a computable number series into small representation and generate them on the go when required. (need lot of improvement to actually apply)