Oleksandr P.

I am a strong advocate for open-source ML tooling: not for ideological reasons, but for practical ones. In the environments I have worked in, vendor lock-in and unpredictable API costs are real operational risks. I have become an expert in the open-source MLOps stack: MLflow for experiment tracking and model registry, Qdrant for vector storage, Airflow for orchestration, and open models like Mistral and Llama for inference.

My MLOps work covers the full lifecycle: DVC for data versioning, automated training pipelines on Kubernetes with GPU scheduling, model serving with BentoML and vLLM, and comprehensive monitoring with Prometheus and Grafana. I have designed systems for clients that can retrain and redeploy models automatically in response to drift alerts with zero manual intervention.

For LLM infrastructure specifically, I specialize in self-hosted deployments: Mistral 7B and Llama models on dedicated GPU instances, quantized for cost efficiency, fronted by vLLM for high-throughput batched inference. I have helped teams reduce their OpenAI spend by 80-90% by transitioning appropriate workloads to self-hosted open models with no perceptible quality degradation.