Solutions · Platform Engineering & SRE
One Helm chart. Resource limits on every container. Standard Kubernetes primitives.
Give product teams a self-service runtime for LangGraph, LangChain, and Deep Agents that passes cluster governance on the first deploy. Zelkor installs with a single Helm command into an existing namespace, sets CPU and memory requests and limits on every container, and works with your current CNI and ingress.
Risks
Where it fails without a runtime boundary
Pods Rejected by Cluster Quotas
Charts that ship containers or init containers without resource limits fail admission on enterprise Kubernetes and OpenShift clusters. How Zelkor stops it: Every container and init container defines CPU and memory requests and limits.
Cluster-Admin Script Sprawl
Installers that rely on imperative shell scripts or try to create namespaces fail under least-privilege CI pipelines. How Zelkor stops it: Deploys purely via Helm into a namespace your cluster admin already created.
Ingress Controller Collisions
AI platforms that demand exclusive ownership of cluster load balancers break existing routing. How Zelkor stops it: Envoy Gateway runs standalone, behind your existing ingress, or attached to a shared cluster gateway.
Accidental Public Exposure of Internal Tools
Misconfigured routes exposing vector databases, guardrails, or MCP servers to the internet. How Zelkor stops it: Only the front door is exposed; internal services never get public routes.
Lateral Movement to State Databases
Unrelated pods on shared clusters connecting to agent checkpoint databases. How Zelkor stops it: Namespace isolation restricts database access to designated agent namespaces.
Proprietary CNI Lock-In
Requiring custom kernel plugins just to restrict pod egress. How Zelkor stops it: Uses standard Kubernetes NetworkPolicy that works across any compliant CNI.
Controls
Compliance & Control Mapping
| Requirement | What Zelkor Does | How You Check It |
|---|
| Helm-Only GitOps Delivery | Installs via helm upgrade --install into an admin-created namespace with no imperative cluster scripts. | Run helm template and verify CI deploy logs. |
| ResourceQuota & LimitRange Compliance | Sets explicit CPU and memory requests and limits on every container and init container. | Deploy into a namespace with an active ResourceQuota. |
| Namespace & Database Isolation | Restricts run-state database ingress to namespaces labeled for agent workloads. | Test database connectivity from an unlabeled namespace. |
| Cluster Monitoring Integration | Includes self-hosted Langfuse for traces plus an optional Prometheus ServiceMonitor for metrics. | Check Prometheus scrape targets and Langfuse UI. |
Topologies
3 Platform Engineering Workflows
Multi-Team Cluster Rollout
Deploy the platform once behind your existing ingress controller and onboard product teams into isolated agent namespaces.
Zero-Boilerplate Agent Onboarding
Application teams bring their LangGraph, LangChain, or Deep Agents container image; the platform wires the AI gateway URL, MCP tools, and tracing automatically.
Dedicated gVisor Node Pool
Schedule the warm sandbox worker pool onto labeled gVisor nodes (or GKE Sandbox / Talos) while running standard agent pods on general-purpose nodes.
Standardize how engineering teams run agents on Kubernetes.