Hardly any eval STRING necessary. You can do entirely without, but it's not practical.

use Carp;
use vars qw(%data);

my %comparison;

sub init_comparison_for {
    local $_ = shift;
    my $op =
        /^string/ ? 'cmp' :
        /^num/    ? '<=>  :
        croak "unknown comparison type $_";

    @comparison{@_} = map {
        eval "sub { \$data{\$a}{$_} $op \$data{\$b}{$_} }"
    } @_;

    @comparison{map "r-$_", @_} = map {
        eval "sub { \$data{\$b}{$_} $op \$data{\$a}{$_} }"
    } @_;
}

sub sortedkeys (\%@) {
    my $hash;
    unless( ($hash = shift) and ('HASH' eq ref $hash) ) {
        croak "not a hashref: $hash";
    }

    local *data = shift;
    my @funcs = @comparison{@_};

    if(my @unknown = grep !defined $func[$_], 0 .. $#_) {
        croak "unknown field name(s): @unknown";
    }

    sort {
        my $r = 0;
        ($r ||= &$_) && last for @funcs;
        $r
    } keys %data;
}
[download]

Use as in

init_comparison_for string => qw(fname lname email address position);
init_comparison_for numeric => qw(id salary zip);

print $employee{$_}->{fname}, " ", $employee{$_}->{lname}
    for sortedkeys %employee, qw(r-salary id);
[download]

This should be improved by taking and returning a hashref in the initialization so there can be multiple different comparison function tables. It might be worthing turning this into an object. You can get even fancier with a tied interface to override the keys behaviour.. the possibilities are endless. (I'm surprised noone has written a module for that yet.)

Makeshifts last the longest.