Neither, really — it's closer to hand-rolled local search.
My first three phases are based on Sugiyama to produce a seed layout (which is broadly what Dagre does). Then I've added my new phases driven by a penalty score that grades candidate layouts using fourteen factors that I devised from my experience laying out graphs manually. The refinement phase is where the bulk of my work is. It applies targeted move types iteratively within a budget, trying to lower the penalty score.
Finally, the result goes through my own embedding and rendering with the theme system on top.
My first three phases are based on Sugiyama to produce a seed layout (which is broadly what Dagre does). Then I've added my new phases driven by a penalty score that grades candidate layouts using fourteen factors that I devised from my experience laying out graphs manually. The refinement phase is where the bulk of my work is. It applies targeted move types iteratively within a budget, trying to lower the penalty score.
Finally, the result goes through my own embedding and rendering with the theme system on top.