Have you tried leveling out the episode in Descript’s export options (you can set a target of -16 LUFS)? If it happens there too, it’s probably an issue with the original audio. Have you also listened between the Auphonic and upload stages? Both of those things could help narrow down where the issue is happening.