We present a novel framework for generating expressive MIDI performances by automatically adding ornaments using PerTok tokenization and a Transformer trained from scratch. In contrast to fine-tuning large language models, our approach employs a domain-specific Transformer designed exclusively for ornamentation, leveraging PerTok's music-aware representation to capture micro-timing and duration nuances critical for expressive performance. The model learns to transform simplified scores into richly ornamented renditions through paired data from the MAESTRO dataset. By training directly on musical data without non-musical pre-training, the system acquires ornamentation patterns in a more stylistically faithful manner. Future work includes comparative evaluation with pre-trained models, exploring alternative tokenization schemes, and conducting systematic evaluation of musicality and stylistic authenticity