Grounded, curious, connected

I'm an AI Quality Engineer and SDET at Apple, where I build the evaluation infrastructure that answers a deceptively hard question: is this LLM system actually getting better, or did the numbers just move?

Seven-plus years of test automation taught me to be suspicious of scores I can't reproduce. So I score deterministically wherever ground truth is computable — tool selection against the logged call trace, values through typed normalizers, hallucination as containment against the retrieved payload — and save the judge model for the one dimension that genuinely needs judgment.

The rest of the job is refusing to let a metric lie. Multi-sample runs so a flaky item isn't mistaken for a regression. Golden answers derived rather than hand-labeled, so the benchmark doesn't rot. Gates tiered so people read them instead of re-running until green.

I grew up in Bishkek, Kyrgyzstan, and the road from the Tien Shan mountains to Silicon Valley shaped how I work: stay grounded, keep learning, and build things that genuinely help people.

"An eval you can't reproduce isn't a measurement. It's an opinion with a number attached."