You can apply the inverse transformation from @auto_tlrc with the methods described in this previous post:
https://afni.nimh.nih.gov/afni/community/board/read.php?1,69100,69588#msg-69588
There are a couple points with your existing processing. First, the application of a separate deobliquing step will cause smoothing from the separate interpolation. You can let align_epi_anat.py take care of the concatenation of the transformations instead, including that obliquity handling, by just using the original input with align_epi_anat.py. Secondly, I think you are using the wrong output from the align_epi_anat.py step, but there may be some additional processing not shown here. The skullstripped output you use for the following step is not the output aligned to the EPI dataset, so that may not be what you want. Lastly, consider auto_warp.py to nonlinearly align your datasets to the template.