Laela Zorana PRO

I write GPU kernels and measure them on real hardware. One global atomic per key took 86.3898 ms on an AMD Instinct MI300X. Privatised into local data share it took 0.1015 ms, which is 851.13 times, on the same card in the same run. The tolerance bound I derived there then held unchanged on an AWS Trainium core. Most of my work sits where two correct answers disagree and something has to decide which one to trust, so I also build the graders that make that call, and the evaluation sets behind them.