Doubled the speed of the implementation:

#! perl -slw
use strict;
use Time::HiRes qw[ time ];

my @c1 = (' ', '0'..'9', 'A'..'Z' );
sub fromB37 {
    my $n = shift;
    my $s = '      ';
    substr( $s, $_, 1, $c1[ $n%37 ] ), $n /= 37 for 0 .. 5;
    $s;
}

my @c2;
$c2[  ord( $c1[ $_ ] ) ] = $_ for 0 .. 36;
sub toB37 {
    my $n = 0;
    $n = $n * 37 + $c2[$_] for reverse unpack 'C*', $_[0];
    $n;
}

my $start = time;
for ( 1 .. 1e6 ) {
    my $s = fromB37( rand 37**6 );
}
printf "fromB37 took %.3f seconds/million\n", time() - $start;

$start = time;
for ( 'AAAAA' .. 'CEXHN' ) {
    my $n = toB37( $_ );
}
printf "toB37 took %.3f second/million\n", time() - $start;

my @data = map int( rand 37**6 ), 1 .. 1e6;
$start = time;
$_ == toB37( fromB37( $_ ) ) or die $_ for @data;
printf "Both ways took %.3f second/million\n", time() - $start;


__END__
C:\test>junk45
fromB37 took 5.309 seconds/million
toB37 took 3.234 second/million
Both ways took 8.953 second/million
[download]