SRE & DevOps engineer · 25 years in IT

Stanislav Nedelchev

I build platforms that survive their own mistakes.

Every platform will see a wrong command, a bad deploy, a leaked key. The question is how far the mistake can spread - and whether the platform, not someone's attention, keeps it small.

I also audit and help build platforms for software teams. Work with me →

Услуги на български →

What I'm building: SafeOps

SafeOps is a course and a working reference platform: a complete Kubernetes platform built and run the way production should be. Every guardrail the course teaches is in the code, enforced by hooks and CI, and tested on a real cluster - a local kind cluster for every lab, and Hetzner Cloud, created, verified and destroyed again, for the cloud track.

Infrastructure as Code

The cluster is code: every change goes through a plan someone has read, against one shared, locked memory of what exists.

  • Terraform
  • Hetzner Cloud
  • AWS
  • GKE
  • Cloudflare
  • k3s
  • kind

GitOps

Git is the only way into the cluster. Flux keeps every environment equal to it - secrets included, encrypted in Git.

  • FluxCD
  • Helm
  • Kustomize
  • SOPS + age

CI/CD

Every change passes the same checks, and production runs exactly the image that was tested in staging - retagged, never rebuilt.

  • GitHub Actions
  • pre-commit
  • CodeRabbit
  • Trivy
  • cosign
  • SBOM
  • provenance

Kubernetes

The cluster itself limits what a mistake can do: who may talk to whom, what a pod may do, how much it may use, and who may change what.

  • Kubernetes
  • NetworkPolicies
  • Pod Security
  • quotas
  • HPA / PDB
  • Kyverno
  • OIDC / RBAC

Kubernetes-ready applications

The platform can only protect an app that keeps its side of the contract: honest probes, a clean shutdown, config from the environment, no root, honest resources, visible signals.

  • Go
  • probes
  • graceful shutdown
  • 12-factor
  • non-root
  • requests / limits

Observability & Reliability

See a problem early, and recover from it - proven by restores and drills, not assumed.

  • Prometheus
  • Grafana
  • OpenTelemetry
  • Uptrace
  • ClickHouse
  • PostgreSQL
  • CloudNativePG
  • backup & restore
  • chaos
  • Flagger

AI-assisted SRE

AI reads the cluster and proposes inside the same guardrails as everyone else; the decisions stay with people.

  • LLM
  • Claude
  • OpenAI
  • k8s-ai-monitor

Four rules for every change

  1. Verify the target before every write.
  2. Plan before apply - and apply only the plan you have read.
  3. One failure domain per change.
  4. Evidence before merge.

Writing

All posts →