Boston, Massachusetts, United States
• Mentored by Professor Brian Kulis
• Developed long context music generation models using Mamba based state space architectures.
• Compared transformer and Mamba pipelines for melody coherence, stability, and scalability.
• Built an experimental workflow using EnCodec tokenization at 24 kHz and 32 kHz with multi-codebook setups.
• Evaluated model quality with CLAP similarity and FAD and used the results to refine architecture choices.
• Prototyped hybrid designs that combine diffusion stages with Mamba sequence modeling.