I build the release platform for Sberbank's Risk division (about 1,500 microservices). Around 150 releases a week now go to production at night on their own. Outside work I find Prometheus alerts that can never fire and fix them in open-source projects.

Open to remote work and relocation · Website · LinkedIn · michael.s.basov@gmail.com

_{Also written as Michael Basov.}

Risk Division · Senior DevOps Engineer · remote · at Sberbank since 2024

• Autonomous deployments. About 150 releases a week go to production unattended, in night and weekend maintenance windows. The system plans each rollout and re-checks the gates before every step. Engineers have almost stopped staying late for releases, and deployment errors went down.
• Release orchestration. Built it and run it (Django, PostgreSQL, async workers). It connects CI with Jira, Nexus and quality gates; coordinating a release went from 2–4 hours to under 10 minutes.
• AI code review. Started and shipped a production reviewer for Bitbucket pull requests: 1,200+ a day, about 30 s per review at the median. It got the internal Engineering Achievement of the Year 2025 award, and I gave a talk about it at an internal conference.
• Kubernetes. I work with 24 of the division's ~100 clusters, dev to production, and run 70+ production services and 6 LLM agents under a 99.9% SLO. Incident response, RCA, automated recovery and graceful degradation.
• RAG assistant for credit managers. In its pilot it closed 30% of routine requests with no human involved.
• Mentoring. Three engineers so far.

Before that (2022–2024): systems engineer at USSC, Yekaterinburg. Enterprise DLP and PAM rollouts, Linux hardening, PostgreSQL, mentoring interns.

Most of these are alert rules that could never fire. I audit widely used Prometheus alert rules for silent or misfiring alerts and fix them upstream with promtool regression tests. AI-assisted, disclosed in each PR. The write-up, Silent or Wrong: Producer-Contract Defects in Shipped Prometheus Alerting Rules, comes out as a preprint in November 2026.

One example: Mimir #16725. MimirRulerTooManyFailedPushes and MimirRulerTooManyFailedQueries matched reason="error", but the ruler records server-side failures as server_error. So a ruler failing 5% of its queries never paged anyone. The selectors now include server_error, keep legacy series and exclude client_error. Synthetic checks: 5% server errors mixed with 20% client errors now fires; client errors alone, or exactly 1% server errors, do not.

Also merged: Russian localization of the Linux security docs in kubernetes/website #55710.

In review: Thanos #9057 · Thanos #9058 · kube-prometheus #2901 · kubernetes-mixin #1306 · Ceph #72211 · postgres_exporter #1397

Claude Code and Codex do the legwork: investigation, patches, tests. I write the plan, review every diff and answer for what gets merged. The AI reviewer above is the production version of the same approach. Upstream PRs say how AI was used and follow each project's AI policy.

Helm · Groovy shared libraries · PromQL / promtool · OpenTelemetry · Dynatrace · ELK · pgvector · RAG · MCP · Qwen

• Certified: AWS Solutions Architect – Associate · AWS CloudOps Engineer – Associate · HashiCorp Terraform Associate
• Studying: BSc Software Engineering at Ural Federal University, part-time, graduating in 2027
• Languages: English B2, Russian native
• Community: judge at CSC Back-to-School Hackathon 2026 (Devpost, 115 submissions); judging CSC Winter Sparkathon, Nov 30 – Dec 6, 2026