Vorelios, a YC Fall 2026 research lab in San Francisco, is launching models for engineering teams that still wait hours or days on a single simulation. The model takes a design's geometry and operating conditions and predicts fluid flow, heat transfer, and structural behavior together, in a fraction of a second, on clean CAD of full-scale designs. Those predictions feed an autonomous search. In the company's first end-to-end runs, the system took a written spec, searched thousands of candidates, and refined the best ones with no person in the loop. The launch video shows that loop on a supersonic jet given one goal, fly at Mach 1.7 while burning as little fuel as possible. The system changes the aircraft, simulates the whole design, and keeps a change only when fuel burn drops. On its results page, Vorelios reports lower pressure and skin-friction error than Transolver-3 and AB-UPT on held-out cases from the NASA Common Research Model, BlendedNet++, and HiLiftAeroML, with a full surface prediction in about 1 to 31 milliseconds on an H100. The founders say those public benchmarks mostly resemble the training data. On families of designs outside that set, earlier models failed, including their previous architecture, and they say closing the rest of that gap is the center of the research. An independent solver, they report, found the first autonomous designs beat a typical candidate, with predictions within a few percent. Daniel McNeil (CEO) and Krish Kapadia (CTO), both 19, started the lab. McNeil studied physics at the University of Florida, and Kapadia did research there in AI and cyber-physical systems. They are generating their own physics data, because they say there is nothing like an internet-scale corpus of simulations to learn from, and they want one model that carries what it learns across fluids, heat, and structures. They are hiring AI researchers and people who work on numerical solvers and CFD.