Data scientist focused on AI safety evaluations — building evals that measure how ML systems actually behave.
Practical AI safety evaluations using inspect_ai (UK AI Safety Institute):
- Position bias detection measured llama2 accuracy drop from 46.7% -> 33.3% under biased prompting conditions
- Toxicity classification with LLM judge architecture on Jigsaw dataset. measured FP/FN rates and failure modes from safety filters
- ReAct agent evaluation on MATH-500. compared no-tools vs naive tool loop vs ReAct; analysed capability attribution when models use external solvers
More details here - treska2000/inspect_ai
8+ years measuring ML systems at Yandex, Avito, Ozon, M.Video.
MSc Mathematics & Mechanics, Moscow State University.
TAIS course (Technical AI Safety), Monoid Center, 2026.
Python inspect_ai Jupyter Ollama LightGBM PySpark Git