status: Running  ·  platform & AI infra @ Aivar Innovation

Manan Jain

$ I

Senior Platform & Cloud Engineer who has run large-scale Kubernetes on AWS — 150+ EKS clusters, ~1,000 nodes, 30K+ pods serving 40+ engineering teams — and built an LLM inference platform on 500+ Spot GPUs at ~60% lower compute cost. Now at Aivar Innovation — EKS, Flyte orchestration, and GPU infrastructure for ML on KhiladiPro.

explore my work → ⬇ download CV
# operator console status: ● building @ Aivar Innovation location: "Bengaluru, IN · IST (GMT+5:30)" response: <24h press: "Ctrl+K to navigate"
⎈ helm status manan · deployed
Manan Jain
$ kubectl get engineer -o wide ✓
$ kubectl top engineer manan --containers
0+
EKS clusters operated
0K+
running pods · ~1,000 nodes
0+
GPUs at ~80% utilization
0+
CI/CD pipelines bulk-upgraded
$ kubectl get projects -A --sort-by=impact

Systems that run entire fleets

Case studies — real problems at 150-cluster, 500-GPU scale, solved with platforms and automation rather than headcount. Each one is a deep dive with architecture and impact.

↗ kubernetes / gpu / flagship

LLM inference on 500+ Spot GPUs

A Kubernetes platform that schedules AI batch workloads on interruptible GPU capacity — checkpointing, retries, and Karpenter elasticity.

▲ ~80% utilization · ~60% compute cost cut
EKSKarpenterFinOps
 ↗ terraform / platform

One module, entire microservice platform

A Terraform module that turns ~50 lines of config into a full ECS Fargate platform — ALB routing, blue/green deploys, a pipeline per service.

▲ weeks of setup → under an hour
TerraformECS FargateBlue/Green
 ↗ python / fleet-scale

Upgrading a 150+ cluster EKS fleet

A 4-phase orchestrator that upgrades EKS clusters fleet-wide — concurrent, cross-account, with automated post-upgrade validation.

▲ quarter-long campaign → supervised batch job
PythonEKSFleet ops
 ↗ argocd / platform

GitOps addon delivery for every cluster

ApplicationSets + multi-source Helm + a per-cluster Git inventory: one commit rolls an addon out to any cluster in the fleet.

▲ per-cluster manual installs → one-line commit
ArgoCDHelmGitOps
 ↗ python / async

Health checks for thousands of services

Async fleet monitoring — concurrent cluster polling, thousands of endpoint probes with aiohttp, Prometheus metrics, actionable reports.

▲ "is the fleet ok?" → 10-minute definitive answer
asyncioPrometheusObservability
 ↗ python / multi-account

500+ pipelines upgraded in parallel

Bulk modernization of CI/CD stacks fleet-wide — idempotent, audited, ring-based rollout, safe to re-run after any partial failure.

▲ hundreds of tickets → one supervised run
boto3ServiceCatalogMulti-account
 ↗ node + react / self-service

Self-service onboarding & multi-tenancy

REST API + React UI that onboards teams end-to-end — Terraform generation, Git automation, notifications — plus multi-tenant app patterns.

▲ engineer-hours per team → a form and a PR
Node.jsReactPlatform
 ↗ python / agentic ai

k8sAgent — agentic SRE copilot

Plan → act → critique until evidence converges. MCP tools, Langfuse-gated evals, RBAC-scoped reads, writes only as GitOps PRs.

▲ evidence-cited answers · bounded loop
KubernetesAgentic AIMCP
 ↗ python / in-cluster monitor

k8lens — in-cluster health radar

Discovers every ingress and service, probes them concurrently with asyncio — because "Running" doesn't mean "up." Prometheus-native.

▲ full-cluster HTTP sweep in seconds
KubernetesasyncioPrometheus
all case studies →
$ kubectl exec -it manan -- /bin/zsh

Talk to my terminal

It's real — type help, kubectl get pods, helm install manan or sudo and see what happens.

manan@prod-cluster: ~ — zsh
connected to prod-cluster. this terminal is real — type a command.
manan@prod-cluster:~$ 
engineer.yaml
apiVersion: devops.manan.dev/v1 kind: PlatformEngineer metadata: name: manan-jain labels: domain: "cloud-infrastructure" scale: "fleet" spec: clusters: 150+ EKS pods: 30K+ gpus: 500+ # ~80% utilized replicas: 1 # unfortunately tolerations: - key: "legacy-infra" effect: "Modernize" restartPolicy: Never # ships reliable systems status: phase: Running conditions: - type: ShippingThings status: "True"
$ kubectl describe services/manan

Different fields, same discipline

Whether you're a startup laying foundations or an enterprise wrangling hundreds of accounts — the playbook is turning recurring problems into systems.

Platform engineering

Reusable IaC modules and golden paths so product teams ship infrastructure in minutes, not sprint cycles.

Kubernetes fleet ops

Upgrades, addon management, and health monitoring designed for hundreds of clusters — not one pet cluster.

CI/CD & release engineering

Blue/green and canary deployments, circuit-breaker rollbacks, zero-downtime releases as the default.

Cloud architecture

Multi-account AWS design, landing zones, migration paths, and cross-account automation that scales with your org.

Reliability & observability

Monitoring and health tooling that surfaces issues before your users (or your pager) do.

Toil elimination

Anything a team does manually more than twice a quarter becomes an automated, audited workflow.

$ cat beyond-the-day-job.md

Writing & building

↗ ✎ blog · 7 posts

Field notes from fleet scale

Seven deep-dives — the 4-layer GPU optimization stack, interruption-native Spot GPUs, fleet upgrade doctrine, inventory-driven GitOps, the grounded agent, and the kro talk.

KubernetesGPU infraGitOps
 ↗ ◆ ventures · builder mindset

Entrepreneurship, off the clock

GlassMic — digitally transforming decorative & packaging glass manufacturing — and AI Smart Insole, an iPhone-scan-to-custom-insole product concept. Complete solutions, not just code.

EntrepreneurshipManufacturingAI product
$ helm install conversation manan/coffee-chats

Have a fleet-sized problem?

I'm heads-down building at Aivar Innovation — not job hunting — but I'm always up for talking infrastructure, GPUs, and agents.

email me → connect on LinkedIn