Abstract: We introduce rolypoly~, the first drum machine for live performance that adapts its microtiming in relation to a human musician. We leverage state-of-the-art work in expressive performance modelling with recurrent nets, towards real-time application on the micro scale. Our models are pretrained on the Groove MIDI Dataset from Magenta, and then fine-tuned iteratively over several duet performances of a new piece. We propose a method for defining training targets based on previous performances, rather than a prior ground truth. The agent is able to adapt to human timing nuances, and can achieve effects such as morphing a rhythm from straight to swing.