Bexultan Kozhageldiyev

Over the past year I shipped a computer-vision pipeline on edge hardware (21/21 QA detection accuracy, ~50% lower CPU), took a stalled vision-language model deployment to production on an RTX 5090 (deploy time 2 h to 15 min), and built the serverless data pipelines and data catalog behind a multi-tenant analytics product on AWS. My thesis was a full ML platform: LSTM demand forecasting and a PPO reinforcement-learning agent, orchestrated in Airflow with an RLHF retraining loop and served through a FastAPI microservice with hot model reload. I work across the whole path: data modelling and ETL, training and evaluation, serving, and the infrastructure underneath. CS degree with honours, two ICPC quarterfinals.