Currently working on better evals for scientific AI at https://teorminimumeval.com

Savelii's open-source AIMO hit top-7% in Olympiad-math solving and documented a ~100× inference speedup on dual T4s; shipped scMagic2, a vector-DB recommender for scRNA-seq tooling; plus prototypes like Turing Hall for interactive math with Lean. He also worked in McKinsey, shipped as early eng in software & hardware startups, and is a cum laude graduate B.S. Physics from MIPT and M.S. from Duke.