Anton Alekseev

Avito · Senior MLOps / SRE · 08.2025 - present

• 15→60% GPU utilization
• >10k RPS
• days→minutes model to production
• 2× fewer pipeline resources

• Raised GPU utilization from ~15% to ~60% on a Kubeflow ML platform (~1,000 monthly active users) running on hundreds of bare-metal Ampere, Hopper and Blackwell GPUs (HGX, NVLink, InfiniBand): Kueue team quotas with checkpointed preemption, topology-aware scheduling, DRA, node health checks and remediation, DCGM dashboards.
• Own the KServe inference platform for ML and LLM end to end, including on-call and SLOs: over 10k RPS across multiple data centers behind a LiteLLM gateway with per-team quotas; cut model time to production from days to minutes.
• Run distributed LLM serving on llm-d with vLLM and SGLang: DP/TP/PP over NVLink and InfiniBand (NCCL), Ray clusters for multi-node inference, disaggregated prefill/decode with a distributed KV cache, KV-cache-aware routing via Gateway API Inference Extension (InferencePool, endpoint picker).
• Tuned vLLM, SGLang and Triton with AIPerf: up to 2.2x throughput and 67% lower inter-token latency through runtime, scheduler and attention-backend tuning; FP8, AWQ and GPTQ quantization, prefix caching, multi-LoRA serving, autoscaling on queue and KV-cache metrics.
• Scaled the platform to ~70k pipeline runs and up to 3k distributed training runs a month across 3,000+ namespaces; wrote Kubeflow plugins and Kubernetes operators in Go; team upstream patches to Kubeflow Pipelines halved pipeline resources.

Yandex Practicum · Program expert, MLOps course (part-time) · 07.2025 - present

• 10 modules
• 5+ authors

• Built a 10-module program, wrote specs and reviewed materials for a team of 5+ authors, designed the course infrastructure and optimized per-student cloud costs.

Selectel · MLOps Engineer · 12.2022 - 08.2025

• 8 h→24 min platform delivery to a client

• Cut ML platform delivery to a client from 8 hours to 24 minutes: on-demand platform on Kubernetes (ClearML, JupyterHub, Keycloak, Istio) with Terraform and GitLab CI instead of manual steps.
• Built a Triton-based inference platform for cloud customers, replacing Seldon Core: canary, autoscaling, scale to zero, weight caching.
• Ran GPU sharing R&D (MIG, time-slicing, MPS, dynamic MIG) and shipped automatic MIG and time-slicing partitioning to the ML platform; launched a CUDA image marketplace.
• Sized infrastructure for LLMs: VRAM math, GPU choice, load testing, vLLM vs SGLang comparison, quantization.

Gazprom Neft · Software Engineer → R&D Engineer → DevOps Engineer · 02.2019 - 12.2022

• Moved on-prem infrastructure in two regions (OpenStack, Proxmox) to IaC with Terraform, Ansible and GitLab CI; set up JupyterHub for the data science team.
• Developed forecasting models (CatBoost) and IIoT equipment diagnostics services (MQTT, Kafka).

Independent consultant · 02.2025 - present

• 20+ consultations

• 20+ one-on-one consultations on ML infrastructure via getMentor: ML platform architecture, GPUs in Kubernetes, sizing infrastructure for LLMs.
• Bank (consulting and custom development): deployed an ML platform on the bank's servers with Helm charts (ClearML, Feast, JupyterHub, KServe) and built an automated training and rollout pipeline for a churn model.