I have a data frame with 2-level indexed columns (sample data below):

  metric         grp        Avg                              P95                     
                           mean     ci_low      ci_up       mean     ci_low      ci_up
0      a     CONTROL   8.202862   8.100596   8.306985  17.122063  16.832979  17.400000
1      a  EXPERIMENT   8.243680   8.141428   8.351050  16.709375  16.383333  17.043500
2      a        diff   0.040819  -0.102122   0.185837  -0.412688  -0.855854   0.050104
3      a    rel_diff   0.005016  -0.012299   0.022944  -0.024022  -0.049377   0.002982
4      b     CONTROL  15.513669  15.316762  15.720606  31.893576  31.357300  32.600350
5      b  EXPERIMENT  15.486001  15.293760  15.672843  31.317389  30.808775  31.839000
6      b        diff  -0.027668  -0.300167   0.251497  -0.576186  -1.450138   0.223063
7      b    rel_diff  -0.001740  -0.019249   0.016368  -0.017967  -0.044785   0.007031  

Result I’d like to produce:

• create a new column in format of mean [cl_low, cl_up] for each level-1 column Avg, P95
• for rel_diff rows, reformat all columns as percentages, e.g.: 0.005016 --> 0.5%.
• for rest rows, round all columns to 3rd digit

What is the pythonic way to do this? Thanks in advance for tips!