- Use to_encoach_dict() for /api/exam/avatars so voiceId, voiceProvider, canvas are included in the response - Set TMPDIR=/tmp in import script for PyTorch compatibility in Docker Made-with: Cursor