Skip to content
kinekube

Platform — the next-gen foundation

Fully managed cloud infrastructure

Production-ready Kubernetes, AI data pipelines, and system telemetry. Hybrid deployments across public and private infrastructure.

Primary deployment targets

  • NutanixPrivate cloud

    Your own data centre, behaving like a cloud region.

  • AWSPublic cloud

    Global reach, managed services, capacity on demand.

Engine room

  • KubernetesThe contract
  • OpenShiftEnterprise K8s
  • Model servingInference
  • Agent runtimeExecution
  • GPU schedulingCapacity
  • MongoDBOperational data
  • Atlas Vector SearchRetrieval
  • PineconeVector store
  • GrafanaOpen stack
  • New RelicManaged APM

What we build

One unified architecture for complex workflows

Four parts to that, and we do all four. Pick the one that matches what is actually hurting — the engineering underneath is the same whichever you start with.

Natively architected for AWS and Nutanix environments. Their roadmap is the platform; knowing the supported path through it is the work.

GOLDEN PATHSDELIVERYOBSERVABILITYKUBERNETESSUBSTRATENUTANIXAWS

The layers

What a platform is, once the demo is over

A working agent in a notebook touches two of these layers. A working agent in production touches all of them, plus the three rails that cut across every one.

Reference architecture

  1. L6

    Experience

    Chat surfaces, internal APIs, workflow triggers, human review queues.

    API gatewayAuth / SSORate & quota policy
  2. L5

    Agent runtime

    Where the agent loop actually executes: planning, tool calls, memory, guardrails, retries.

    OrchestratorTool sandboxGuardrailsEval harness
  3. L4

    Model serving

    Inference endpoints with real batching, routing and fallbacks — self-hosted, managed, or both.

    vLLM / TritonBedrockRouter & fallbackKV cache
  4. L3

    Data & retrieval

    The context layer. Operational state, embeddings, documents and the pipelines that keep them fresh.

    MongoDBAtlas Vector SearchPineconeIngest pipelines
  5. L2

    Platform

    The contract that makes everything above portable: clusters, scheduling, delivery, secrets, identity.

    KubernetesOpenShiftGPU schedulingArgo CD
  6. L1

    Substrate

    Compute, network and storage — a cloud region, a rack in your building, or a mix of the two.

    AWSNutanixGPU nodesNetwork fabric

Spanning every layer

Observability

Traces from the ingress through every tool call and model invocation. Token spend, latency and failure class attributed per agent.

OpenTelemetryNew RelicGrafana LGTM

Governance

Who can call which model with what data, proven after the fact. Prompt and output retention with a defensible policy.

Policy engineAudit trailData residency

Cost control

Unit economics per conversation, per tenant, per model. Budgets that alert before the invoice does.

ShowbackCapacity planSpot & reservations

Placement

Three honest answers to “where should this run?”

There is no universally right substrate — only a right one for your data, your demand curve and your balance sheet. Here is how we advise on each.

On public cloud

When speed and elasticity win

You need capacity this quarter, your data has no residency constraint, and demand is spiky. We build a landing zone with the guardrails already in place, then run inference on EKS with GPU node pools that scale to zero.

Choose it when

  • Bursty or unpredictable inference demand
  • Small platform team, no data-centre footprint
  • Managed model endpoints acceptable for most traffic

Budget for

  • Per-token cost becomes the dominant line item at scale
  • GPU availability is a scheduling problem, not a purchasing one
  • Data egress quietly funds someone else's roadmap

On-premises

When the data cannot leave

Regulated data, sovereignty requirements, or an inference bill that now exceeds the cost of owning the hardware. We build on Nutanix so the private footprint still behaves like a cloud region — self-service, API-driven, and automated with the same Terraform.

Choose it when

  • Regulated or sovereign data that cannot cross a border
  • Sustained, predictable inference volume
  • Existing data-centre capacity and a capex-friendly balance sheet

Budget for

  • Capacity planning becomes a real discipline again
  • GPU driver, firmware and kernel lifecycles need an owner
  • Physical redundancy has to be designed, not assumed

Hybrid

Hybrid deployments across public and private infrastructure

The most common outcome. Sensitive retrieval and fine-tuning stay next to the data on private infrastructure; overflow inference and experimentation burst to public cloud. Identity, policy, delivery and telemetry stay unified so it is one platform, not two.

Choose it when

  • Mixed data classifications across the same product
  • Owned baseline capacity with cloud for the peaks
  • A credible negotiating position with every vendor

Budget for

  • Network path between substrates must be designed first
  • Two failure domains means two on-call runbooks
  • Identity and secrets need a single authority, not a copy

Part by part

Where our hands actually are

Twelve technologies, grouped by the part of the platform they belong to, and the specific work we do inside each.

Platform

The substrate and the contract on top of it. A rack in your building and a cloud region should present the same interface to the teams using them — that is what makes placement a commercial decision rather than a technical one.

Nutanix

Private cloud

When data cannot leave the building, Nutanix is where we land it. We design AHV clusters, Nutanix Kubernetes Platform footprints and Objects storage tiers that behave like a cloud region — with the same automation contract as AWS.

Engagements

  • AHV cluster sizing for GPU and high-memory inference nodes
  • Nutanix Kubernetes Platform cluster lifecycle
  • Nutanix Objects as an S3-compatible model and artifact store
  • Prism Central automation via Terraform and the v3/v4 APIs
  • Stretch clusters and Metro availability for stateful workloads

Kubernetes

The contract

The one interface that makes private and public substrates look the same. We build clusters, write operators, and define the golden paths that let application teams ship without filing tickets.

Engagements

  • Cluster builds: kubeadm, EKS, Nutanix Kubernetes Platform
  • Custom operators and CRDs for domain-specific automation
  • GitOps delivery with Argo CD and progressive rollout policy
  • GPU scheduling, device plugins, MIG partitioning and node pools
  • Upgrade trains, admission policy and multi-tenancy isolation

OpenShift

Enterprise K8s

Where an organisation has standardised on Red Hat and the platform has to satisfy a security team as well as a delivery team. We work with the opinionated parts rather than around them, so support remains a straight line.

Engagements

  • Cluster installation and Operator Lifecycle Manager workflows
  • Security Context Constraints and admission policy that survive audit
  • OpenShift GitOps and Pipelines as the delivery path
  • GPU Operator and Node Feature Discovery for inference nodes
  • Migration paths between OpenShift and upstream Kubernetes

AWS

Public cloud

The default landing zone for burst, global reach and managed services. We build multi-account organisations with Control Tower guardrails, EKS platforms, and GPU capacity strategies that do not quietly bankrupt you.

Engagements

  • Landing zones: Organizations, SCPs, network segmentation, IAM boundaries
  • EKS platform builds with Karpenter, Gateway API and IRSA/Pod Identity
  • GPU capacity planning across On-Demand, Spot and Capacity Blocks
  • Bedrock and SageMaker integration for hybrid inference paths
  • FinOps: tagging contracts, budgets, and per-workload unit economics

AI services

The machinery between a prompt and an answer: inference endpoints that batch properly, GPUs that are not idling, and an agent loop you can reproduce when it misbehaves.

Model serving

Inference

A model behind an endpoint is not a serving layer. We build the routing, batching and fallback behaviour that decides whether your latency budget holds when traffic arrives in a burst.

Engagements

  • vLLM, TGI and Triton deployments with continuous batching
  • KV cache reuse and prefix caching for repeated context
  • Model routing, version pinning and graceful fallback chains
  • Managed endpoints blended with self-hosted for cost control
  • Load testing against your real traffic shape, not a synthetic one

Agent runtime

Execution

Where the agent loop actually executes. Tool calls have side effects, models time out and plans go wrong — the runtime is what makes those events survivable and, more importantly, explicable afterwards.

Engagements

  • Tool sandboxing with timeout, retry and idempotency policy
  • Guardrails on input and output, including PII handling
  • Evaluation suites that run on every prompt or model change
  • Human review queues for decisions that should not be autonomous
  • A reproducible trace for every run, so a bad answer has a cause

GPU scheduling

Capacity

Accelerators are the most expensive idle resource in the building. We make them busy without letting an experiment take production down with it.

Engagements

  • Node pools, device plugins, MIG partitioning and time-slicing
  • Queueing and preemption rules that protect production workloads
  • Driver, firmware and kernel lifecycle ownership on-premises
  • Scale-to-zero topologies for bursty or batch inference
  • Capacity models that compare owning against renting honestly

AI production data

Retrieval is where most AI systems quietly fail. Operational state, embeddings and the pipelines that keep them fresh, built to survive a corpus that grows and a schema that changes.

MongoDB

Operational data

The operational database behind most agent systems we build — documents for state, and a familiar query surface for the application team. We run it self-managed with Ops Manager and the Kubernetes Operator, or on Atlas, depending on where your data has to live.

Engagements

  • Ops Manager deployment, backup daemons and oplog store design
  • MongoDB Enterprise Kubernetes Operator for replica sets and sharding
  • Schema, index and shard-key review before you feel the pain
  • Zero-downtime version upgrades and cross-environment migrations
  • Restore drills, because an untested backup is a rumour

Pinecone

Vector store

Where retrieval is the dominant workload and the team would rather not operate the index themselves. We treat it as one credible option among several, and will say when co-locating vectors with your operational data is the better answer.

Engagements

  • Index topology, namespace design and metadata filtering strategy
  • Ingest and re-embedding pipelines with backfill and replay
  • Hybrid and sparse-dense retrieval tuning against an eval set
  • Cost modelling per namespace as the corpus and traffic grow
  • Migration paths to and from self-hosted or Atlas-based retrieval

Observability

Traces from the ingress through every tool call and model invocation, with latency, spend and failure class attributed to something you can act on.

Grafana

Open stack

For teams that want to own their telemetry. We build the LGTM stack — Loki, Grafana, Tempo, Mimir — on your own infrastructure, with retention and cardinality budgets set deliberately instead of discovered on an invoice.

Engagements

  • Mimir, Loki and Tempo deployed for multi-tenant, long retention
  • Prometheus and OpenTelemetry collector topologies at cluster scale
  • Cardinality budgets and recording rules that keep queries fast
  • Golden dashboards per service, generated from a shared library
  • Alertmanager routing tied to real ownership, not a shared inbox

New Relic

Managed APM

Where teams want one pane of glass with APM, logs and traces already correlated. We instrument the whole path — from the ingress to the model call — so an agent's latency has an owner, not a theory.

Engagements

  • OpenTelemetry pipelines with consistent resource attributes
  • Kubernetes cluster and pod-level instrumentation
  • LLM and agent observability: token cost, tool latency, failure class
  • SLO definition, alert quality reviews and on-call noise reduction
  • Ingest governance so telemetry spend stays proportional to value

Cross-cutting

The work that belongs to every part

Some of it belongs to no single part, which is exactly why it is the work most often missing from a proposal. We do it on every engagement, whichever part brought you here.

Foundation

Landing zones & provisioning

Accounts, networks, identity boundaries and DNS built as Terraform modules — so environment number four takes an afternoon rather than a fortnight.

Delivery

CI/CD & GitOps

Pipelines that build, scan, sign and promote artefacts, with Argo CD reconciling every environment from Git. Rollback is a revert, not an incident.

Assurance

Security & compliance

Secrets management, PKI and rotation, least-privilege IAM, image provenance, network policy and the evidence trail your SOC 2 or ISO auditor will ask for.

Economics

Cost & capacity

A tagging contract that survives contact with reality, per-team showback, rightsizing, commitment planning and headroom modelling before the traffic arrives.

Continuity

Resilience & DR

Failure domains mapped, recovery objectives agreed and actually exercised. Game days where we break things on purpose while everyone is awake.

Operations

Run & on-call

Patching, upgrades, capacity changes and incident response against agreed service levels, with post-incident reviews that produce commits rather than slides.

How we use vendor depth

Depth should reduce your risk, not ours

  1. 01

    The vendor already solved it

    Every platform we work with has a supported path for the thing you are about to build by hand. Knowing that path is most of the job, and it is why our builds survive support tickets.

  2. 02

    Everything is code

    If it was clicked into existence, it does not exist. Terraform for infrastructure, Helm and Argo CD for workloads, OpenTelemetry for signals — one automation contract, whichever face you came for.

  3. 03

    Substitution, not lock-in

    We deliberately keep a credible alternative to each choice — Nutanix against AWS, Grafana against New Relic, Atlas Vector Search against Pinecone. Two workable answers is what gives you commercial leverage.

  4. 04

    No black boxes

    You get the repositories, the diagrams, the runbooks and the reasoning. We size entitlements against the workload you actually have, and we will say so when a cheaper tier does the job.

Or have us build the whole thing

Architecture call

Bring the problem. We will bring the architecture.

Workload shape, data classification and where it has to live are enough for us to sketch a design and an honest run cost on the first call.

Book an architecture callSee the software factory