At Polymath Labs, I built RL environments and graders for evaluating frontier LLMs on software-engineering tasks. After finding models that passed by editing tests, I added protected-file checks and hidden test sets to every grader.
I also consulted 4 B2B clients on agent evaluation and translated their requirements into environment and grader specs. I ran episodes in parallel over NATS and Redis queues, with environments in Docker and Kubernetes.
At Nomad Venture Studio, I built an LLM pipeline for App Store and Trustpilot reviews, covering topic classification and summaries. I audited ROI reporting on 3,500+ ad creatives and found a metric error that inflated ROI by up to 19 pp; I also built revenue anomaly monitoring that caught an unreported 24.7% drop.
At ASP-LAB, Nazarbayev University, I owned ML pipelines from annotation through training and validation, and compared inference speed and accuracy before and after ONNX export and INT8 quantization. I also owned the energy-sector forecast in Kazakhstan's national LEAP emissions model at Climate Change Coordination Centre.

