Iaroslav Aksenkin

I’m a Machine Learning Engineer with 4+ years of experience building models end-to-end—from data infrastructure and evaluation tooling for real-time detection to training and production deployment. My work consistently focuses on measurable model quality and fast iteration through solid tooling.

In my current role, I built the core training-data pipeline for real-time detection systems, pulling post-review field data into datasets reusable for training and evaluation. This cut per-iteration dataset setup from about half a day to roughly an hour (~4x faster), and I also helped train production detection models, reaching 0.65 mAP on held-out validation.

I design evaluation pipelines that catch failure modes early: I built internal evaluation tooling that runs pipeline components against curated, domain-specific datasets to surface per-domain issues before they reach the detector. I’ve also strengthened training reliability with experiment tracking (MLflow), flexible class-mapping modules, and automated, reproducible CI builds for containerized model packaging.

Beyond computer vision, I’ve delivered applied LLM solutions for real-world services, including multi-provider connectors, prompt engineering, agent-tool connectors, and an evaluation pipeline using GEval (improving project-level GEval from 0.30 to 0.83). I’m comfortable working across LLM integration, model training, and MLOps, including AWS-based services and end-to-end automation.