Anton Alekseev · ML infrastructure
ML and inference platforms on Kubernetes: GPU sharing, scheduling, inference autoscaling.
Anton Alekseev, MLOps engineer. I build an ML platform for 1000+ users and ML and LLM inference at a large classifieds marketplace; before that, an ML platform and inference at a cloud provider. Program expert of the MLOps course at Yandex Practicum, speaker at HighLoad++, DevOops and PyCon.
You can get a model running with docker run in an evening. Then it has to survive real load, share GPUs with other teams, get updates without downtime and not burn through the budget. That part is what I do.
DAY 0 · DESIGN
DAY 1 · LAUNCH
DAY 2 · OPERATIONS
Need your own LLM at your company? Choosing infrastructure and turnkey setup →
5%Average GPU utilization in production Kubernetes clusters, according to the Cast AI 2026 report. An idle card costs as much as a busy one.cast.ai ↗
What I built and what came out of it. Before and after numbers from real projects.
large classifieds marketplaceML platform
15%→~60%
GPU utilization: from a DevBox per person to one platform
~1000 MAU~70k runs/monthKubeflow · Kueue · Istio
large classifieds marketplaceinference
days→minutes
from a model in the registry to production, instead of a web service in the PaaS mixed with business logic
>10k RPSgeo-distributed clusterprefill/decode
cloud providerML platform
8 h→24 min
to deploy the platform for a client
bankNDA
Helm → prod
a platform on the bank's servers and a churn model with automated training and rollout
Yandex Practicumeducation
10 modules
program expert for every module: responsible for the curriculum
getMentormentoring
Anton does not just know the subject - he is a practicing professional, and his answers come from hands-on experience.
George
Eleven steps from the first GPU in a cluster to a platform where teams train and deploy models on their own.
Open →career trackA skill map at the intersection of ML, development and operations. Rate yourself and download a plan of what to learn.
Open →I'm breaking down a CNCF overview on fast vLLM and SGLang startup: which techniques work (weight caches, snapshots, CUDA IPC), which are still on the roadmap, and which startup stages still cost you money.
Inference · LLMs and hardware
I updated the interactive MLOps engineer roadmap: now you can build a personal learning track by Junior, Middle, and Senior levels, get project recommendations, and export your plan as JSON or Markdown.
ML platforms · Career and community
NVIDIA released the full Run:ai platform source code under Apache 2.0. I'll walk you through what's inside: control plane, KAI scheduler with gang scheduling, GPU fractioning with hard memory isolation, and multi-cluster management tools.
Scheduling and utilization · ML platforms
infra.conf 2026 (Yandex Infrastructure), Moscow
MLechny Put 2026 (Selectel ML meetup), Moscow
ITMO x IT Infrastructure Club (podcast)
self x Zvuk career meetup, Moscow
The easiest way is to message me on Telegram.
