Note that you would not have an encoding for negative numbers then, and you would have to use the alternative base64 table,

  +/0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz

as the normal one is not order-preserving. Otherwise yeah, you could probably do that--and of course there are many ways to patch what you've got to handle negatives, e.g. by noticing that it is slightly wasteful (the base64 prefix ++++++++ is never used) or by using a leading character lower than + (so - is out but ! for example is available) and then encoding the rest of the string in an "ascii-reversed" way.

However I must object to the idea that "real hardware" cannot handle the petabyte scale. The in-article recursive approach of using the algorithm:

    function encode(number) {
        //calculate some base number_string
        return number < NUM_THRESHOLD?
            number_string :
            sigil + encode(number_string.length) + opt_divider + number_string;
    }

is not a particularly bad one. The only thing that maybe I'd balk at is that the given encoding requires a human editor to use an ASCII table to lookup their string lengths, which is clearly undesirable... one might instead go for the most naive case where NUM_THRESHOLD is simply 10 and have the number 123,456,789,012,345,678,901 be represented as:

    ~~2 21 123456789012345678901

so that it's clearly "there are going to be 2 length strings" (hence two tildes) -- the first always has length 1, it specifies "the next has length 2" and then "the next has length 21."

This is also pretty easy to parse and one can use a leading - sign to encode negative numbers, for example.