Co-developed MIDIGPT with researcher Peter Bloem at the VU Amsterdam Network Institute — a Transformer-based music generation system that tackles two fundamental limitations in AI music: the lack of micro-level expressivity and the bias toward Western musical structures. I designed custom tensor representations that encode the timing, velocity, and structural nuances that make music feel human — elements that standard MIDI tokenization discards. The research was supported by an NWO small grant (~€30k) and trained at scale using distributed workloads across H100 GPUs on the Dutch national supercomputer Snellius. Part of this research has been concluded in my Master's thesis, which can be found here.