Hi, I’m Rosalia. I love pushing systems to the edge of what is possible—testing new algorithms, exploiting strange compiler tricks, then reducing the result to a small algebra.
Moonlight Triangulation is my composable Haskell port of Spade’s Delaunay engine. The meshes it produces can be joined, intersected, subtracted, refined, and folded as ordinary values. Across stable benchmarks, Moonlight is on par with or faster than Spade, and at scale the lead can become dramatic: in one 25k/100k removal benchmark, Moonlight finished in 84.668 ms versus Spade’s 525.134 ms—more than six times faster. Another project, Nebula, is an equality-saturation engine: it explores equivalent ways to run an AI model and selects the fastest legal program for each kind of hardware. On a nine-target benchmark, one shared search finishes in 22.6s; egglog, another engine that searches equivalent programs, takes 69.7s to search each hardware target separately.
My current undertaking is Nebula’s megakernel search: deriving hardware-aware, vendor-legal Qwen decoder programs, proving them equivalent, and handing the survivors to real GPUs for judgment.
I came to compilers through product engineering rather than a straight research path. In tech contracting, I traced production failures, made interfaces load faster, and built safeguards for AI agents that could silently stall. I later created and sold Gator, a tool that reviews entire codebases, and worked independently across frontend, backend, telemetry, and scaling. That range keeps my research grounded: however unusual the machinery gets, it still has to be fast, understandable, and useful to the people relying on it.
Outside of work, I like making physical things too. I hand-built the pair of Dactyl keyboards on my desk—split, sculpted keyboards shaped around the hands instead of a rectangle. Apparently, an unreasonable amount of soldering and iteration is what I consider relaxing.
