Platform engineering01
Make the right path the easiest path.
Kubernetes foundations, GitOps delivery, and policy boundaries as developer-facing primitives.
Germany
Andrey Lesnikov
Infrastructure
that runs itself.
Senior DevOps & AI Platform Engineer · Kubernetes · AI Infrastructure
Cloud-native platforms, governed AI infrastructure and resilient systems — designed to stay calm when production gets loud.
Control plane
one plane · three loops
LOOP 01 · REHEARSALapproved
LOOP 02 · RUNTIMErouted
LOOP 03 · HEALINGrecovered
01
Engineering principle
Complexity belongs behind the interface.
The platform can be sophisticated underneath. The experience should still feel calm, legible and predictable — for developers, operators and the business.
Areas of work
Kubernetes platforms, governed AI, and proved recovery.
AI infrastructure02
Govern every request, not just the model.
Routing, identity, MCP policy, inference and observability as one execution plane.
Reliability systems03
Restore should be proved, not assumed.
Recovery contracts with a binary PROVED / UNPROVED verdict — protect, validate, restore, prove.
02
Selected systems
Nine featured repositories. Nine different kinds of proof.
Three flagship systems and six more repositories — each verified in CI on kind, minikube, or Compose labs.
01 / Architecture Rehearsal
Test the architecture
before production does.
Deterministic failure rehearsal for architectural assumptions, policy gates and evidence chains.
View case ↗REHEARSAL / 04approved
argo-presync · netpol-cut
BLOCK
02 / TwinOps Control Plane
Treat industrial twins
like infrastructure.
GitOps control plane for industrial digital twins — compose, detect drift, reconcile, replay incidents.
View case ↗TWINOPS / CONTROLsynced
kit streamlive
gitopssynced
03 / AI Runtime Platform
Every AI request
passes through policy.
OpenAI-compatible gateway, MCP governance, KServe/vLLM, KEDA, OIDC, Redis and OpenTelemetry.
View case ↗AI RUNTIME / GATEpolicy on
identityoidc
e2e21.5s
verdictallow
Govern private AI as a platform product.
Policy, identity, audit, MCP, FinOps and SLOs as one operating model.
View case ↗Detect. Recover. Verify. Then trust.
Failure injection and bounded remediation loops on the platform.
View case ↗Argo CD vs Flux as an operating model duel.
N=3 CI benchmark of ready, self-heal and recovery times on kind.
View case ↗Don't count CVEs — predict where they can reach.
SBOM, VEX, runtime presence → ALLOW / WARN / QUARANTINE / BLOCK.
View case ↗Restore proved before the incident.
PROTECT → VALIDATE → RESTORE → PROVE with binary PROVED / UNPROVED.
View case ↗Fail-closed GPU admission on Kueue.
AdmissionCheck ALLOW / HOLD / DENY — 13 unit scenarios, 8 kind e2e.
View case ↗GitHub ↗Book
Prove It Before Production
Evidence-Driven Infrastructure Engineering for Kubernetes and AI
How do you know that your infrastructure change is actually safe?
Most platform books show how to assemble Kubernetes, GitOps, portals and pipelines. This book asks the harder question: when terraform plan is clean, CI is green and GitOps reports Synced, what evidence do you actually have?
Recognition
Cloud Native Rockstars 2026
Professional recognition is useful when it points back to real systems, real constraints and work that holds up under scrutiny.
Open proof page ↗
Hackathons
Two public builds. Both let the model propose, never decide.
Public builds with what they show, what was measured on their own corpus, and what they do not claim.
See the builds ↗
Nebius × NVIDIA · Devpost · 30 October 2026
BlastGuard AI
A secure patch can still break production. NVIDIA Nemotron only proposes the YAML fix; an Intent Preservation Gate, which does not depend on the model, decides.
Read the build ↗
Vultr: Agent Rush · Blast Radius Zero · 5 November 2026
Airlock
An execution firewall for AI agents. The model only proposes actions; a deterministic policy engine decides allow or deny, with default deny.
Read the build ↗
Available for meaningful infrastructure conversations