Engineering reliable systems at scale

Paras Bhanderi — Senior Site Reliability Engineer

Senior Site Reliability Engineer

Platform Engineering·AI Infrastructure·Distributed Systems

I build resilient cloud platforms, automate infrastructure end to end, and engineer the systems modern AI workloads run on.

Explore my work GitHub LinkedIn Résumé

Currently: Senior SRE @ Shell
Location: Bangalore, India
Certified: CKA · Terraform · AZ-400 · AZ-104 · AI-102

scroll

01Identity

Reliability engineering, cloud platforms and the infrastructure AI runs on.

7+ years in SRE, DevOps and infrastructure software engineering — building the Python automation, Git-driven CI/CD and release tooling that engineering teams depend on.

At Shell I own build, release and notification tooling for enterprise golden-image pipelines, turning recurring manual infrastructure work into automated, observable services. Before that: event-driven self-healing automation and a GitOps CI/CD transformation at Gore Mutual, and reproducible multi-cloud infrastructure and Kubernetes disaster recovery at Kompusys.

More recently I’ve been applying that foundation to AI infrastructure: an Azure AI Foundry reporting agent at Shell, MCP-based systems with tool-calling agents, self-hosted model serving with vLLM on Kubernetes, RAG and LLMOps.

01
Automate the operational surface
If a human does it twice, it becomes code — Terraform modules, pipelines, runbooks and self-healing workflows that remove toil instead of documenting it.
02
Design for failure
Nodes die, deploys break, regions degrade. Health checks, autoscaling, ordered rollouts, backups and disaster recovery turn outages into non-events.
03
Make infrastructure observable
You can’t operate what you can’t see. SLIs, SLOs, metrics, logs and alerts that shorten the distance from signal to root cause.

paras@platform: ~

zsh

➜~

run

principles/platform.py

python

1class Platform:2    reliability = "designed"3    automation = "default"4    observability = "mandatory"

02Impact

Production outcomes, measured.

Every number below comes from platforms operated in production, and each is tied to the work that produced it.

impact.metricssource: résumé · 2019 → present

release.frequency▲ increase
30%
Release frequency
Owning CI/CD, release and deployment for golden container and machine image pipelines.
Shell
deploy.failures▼ reduction
35%
Deployment failures
Same image pipelines on GitHub Actions, Azure DevOps and Packer — fewer failed releases.
Shell
deploy.manual_effort▼ reduction
60%
Manual deployment effort
Terraform and Packer IaC for golden image builds and multi-region replication across AWS and Azure.
Shell
incident.mttr▼ reduction
25%
Mean time to recovery
Pipelines and platform services instrumented with Prometheus, Grafana and Dynatrace.
Shell
vm.remediation05
26
Production VMs remediated
Traced to a misconfigured Windows Server 2022 golden image; snapshot, isolated validation, batched rollout.
Shell
release.cycle_time▼ reduction
50%
Release cycle time
CI/CD transformation to GitHub Actions, Argo CD and GitOps with Git-driven releases.
Gore Mutual
ops.toil▼ reduction
35%
Operational toil
Event-driven Python self-healing automation triggered by monitoring alerts.
Gore Mutual
observability.coverage08
50+
Microservices observed
Standardized dashboards and alerting on AWS and Azure; troubleshooting time down 30%.
Gore Mutual
platform.uptime09
99.9%
Platform uptime
Disaster recovery for Kubernetes-hosted SQL and NoSQL workloads, with backup and availability work.
Kompusys

03Experience

Operating cloud infrastructure in production since 2019.

From root-cause analysis on enterprise fleets to GitOps delivery, golden image pipelines, self-healing automation and AI agents in production workflows.

Apr 2026 — Present
current role
Apr 2026 — Present
Shell
HEAD → shell
Senior Site Reliability Engineer
Golden container and machine image pipelines, multi-cloud IaC, observability and AI agents across AWS and Azure.
- Own CI/CD, release and deployment for golden container and machine image pipelines in GitHub Actions, Azure DevOps and Packer+30% release frequency · −35% deployment failures
- Secured build alerting for 3 production image pipelines through Microsoft’s O365 connector retirement as sole engineer — migrated webhook integrations to Power Automate and rewrote the shared Python notification action3 pipelines kept alerting
- Terraform and Packer infrastructure as code for golden image builds and multi-region image replication across AWS and Azure−60% manual deployment effort
- Python, Bash and PowerShell tooling for recurring infrastructure operations−50% manual intervention
- Instrumented pipelines and platform services with Prometheus, Grafana and Dynatrace−25% MTTR
- Traced vulnerabilities to a misconfigured Windows Server 2022 golden image (disk controller, CIS hardening, accelerated networking) and led a snapshot, isolated validation and batched rollout26 production VMs remediated
- Designed a weekly 2x2x2 reporting agent in Azure AI Foundry — Azure Logic Apps gather Azure DevOps work items, comments and history, and AI drafts route through Microsoft Teams for employee review and approvalAI agent · human-in-the-loop
- GitHub Actions
- Azure DevOps
- Packer
- Terraform
- AWS
- Azure
- Windows Server
- Prometheus
- Grafana
- Dynatrace
- Power Automate
- Azure AI Foundry
- Logic Apps
- Python
- Bash
- PowerShell
Shell · engineering ecosystemtools named in résumé
GitHub Actions and Azure DevOps drive Packer and Terraform golden image builds replicated across AWS and Azure regions and observed with Prometheus, Grafana and Dynatrace; pipeline events alert through Power Automate, and an Azure AI Foundry agent fed by Logic Apps drafts reports for review in Microsoft Teams.
Nov 2022 — Mar 2026
3 yr 5 mo
Nov 2022 — Mar 2026 · 3 yr 5 mo
Gore Mutual Insurance
Site Reliability / DevOps Engineer
GitOps delivery, multi-cloud provisioning, observability and self-healing automation.
- Event-driven Python self-healing automation triggered by monitoring alerts across multi-cloud production systems−35% operational toil
- Led a CI/CD transformation to GitHub Actions, Argo CD and GitOps with container-based packaging and Git-driven releases, strengthening compliance and security posture−50% release cycle time
- Led Linux, networking and performance root cause analysis across cloud-native stacks−25% MTTR
- Built standardized dashboards and alerting across 50+ AWS and Azure microservices, speeding incident detection−30% troubleshooting time
- Standardized multi-cloud provisioning with reusable Terraform modules and Ansible roles−40% manual operational effort
- GitHub Actions
- Argo CD
- GitOps
- Containers
- Terraform
- Ansible
- AWS
- Azure
- Linux
- Python
Sep 2020 — Oct 2022
2 yr 2 mo
Sep 2020 — Oct 2022 · 2 yr 2 mo
Kompusys Consultant Inc.
Platform Engineer
Reliability of cloud platforms serving thousands of users, Kubernetes data workloads and reproducible IaC.
- Resolved database, authentication and performance bottlenecks on Linux for cloud platforms serving thousands of users−35% incidents & MTTR
- Spearheaded disaster recovery strategies for Kubernetes-hosted SQL and NoSQL workloads through robust backup and availability enhancements99.9% uptime
- Infrastructure-as-code frameworks that made AWS and Azure deployments reproducible−50% environment setup time
- Automated container and platform operations with Python and Bash, improving deployment consistency−30% operational overhead
- Kubernetes
- Infrastructure as Code
- AWS
- Azure
- Linux
- Python
- Bash
- Backup & DR
Dec 2019 — Aug 2020
9 mo
Dec 2019 — Aug 2020 · 9 mo
Apple Inc.
via Kelly Services
Technical Support Engineer
Production support and root cause analysis across enterprise Linux and macOS environments.
- Resolved escalated networking and operating system issues across enterprise Linux and macOS environments−30% downtime
- Engineered disaster recovery solutions for Kubernetes workloads, minimizing potential data loss during critical system failures99.9% uptime
- Performed root cause analysis for high-severity incidents across systems, networks and hardware, working with engineering and security teams on device policies and platform stability
- Linux
- macOS
- Networking
- Kubernetes
- RCA
- Backup & DR

04Platform Engineering

Kubernetes, GitOps and IaC — delivered as a platform, not a pile of scripts.

Git is the source of truth, reconciliation is continuous, and every environment is reproducible. The patterns I apply in production, packaged as a reference platform.

PROJECTeks-gitops-platform

GitOps Platform Engineering

EKS · Kubernetes · Argo CD · Terraform

View on GitHub

pipeline · main → ekssimulated run

01/08

›devdeveloper opens pull request

environment promotion

01DEV○ Pending
02STAGING○ Pending
03PRODUCTION○ Pending

A GitOps platform built on Argo CD and declarative workflows, with modular infrastructure and application patterns that scale deployments across development, staging and production.

better deployment consistency: 50%▲better deployment consistency
fewer manual releases: 40%▼fewer manual releases
faster service onboarding: 30%▲faster service onboarding
fewer deployment failures: 35%▼fewer deployment failures

Declarative infrastructure
Continuous reconciliation
Self-healing deployments
Environment consistency
Application onboarding patterns

infra/eks.tf

illustrative

1resource "aws_eks_cluster" "platform" {2  name     = "ai-platform"3  role_arn = aws_iam_role.cluster.arn4 5  vpc_config {6    subnet_ids = module.vpc.private_subnets7  }8}

gitops/apps/sample-api/argocd-app.yaml

from repo

1apiVersion: argoproj.io/v1alpha12kind: Application3metadata:4  name: sample-api5  namespace: argocd6spec:7  syncPolicy:8    automated:9      prune: true10      selfHeal: true

05AI Infrastructure

From cloud platforms → AI infrastructure.

The primitives that keep production platforms reliable — scheduling, isolation, safe rollouts, identity and observability — are exactly what AI workloads depend on. This is the direction of my specialization, labelled honestly by depth.

Operated in production
Built & shipped

01
Cloud Infrastructure
Operated in production
- AWS
- Azure
- GCP
- VPC / VNet
- DNS
- Routing
- Load Balancing
↳ Multi-cloud environments at Shell, Gore Mutual and Kompusys
02
Platform Engineering
Operated in production
- Terraform
- Packer
- Ansible
- GitHub Actions
- Azure DevOps
- Argo CD
↳ Golden image pipelines at Shell · GitOps transformation at Gore Mutual
03
Kubernetes
Operated in production
- EKS
- AKS
- GKE
- Helm
- Kustomize
- Crossplane
- HPA
↳ Kubernetes data platforms with DR · MCP Platform on EKS · CKA certified
04
Distributed Systems
Operated in production
- Microservices
- Prometheus
- Grafana
- Dynatrace
- RCA
↳ Dashboards and alerting across 50+ microservices
05
Model Serving
Built & shipped
- vLLM
- Gemma 3
- OpenAI-compatible API
- MLOps
↳ vLLM inference on Kubernetes · Cost Spike Detector · ML & Deep Learning Specializations
06
LLM Infrastructure
Built & shipped
- Azure AI Foundry
- AWS Bedrock
- RAG
- LangChain
- LLMOps
↳ Azure AI Foundry reporting agent at Shell · AI-102 Azure AI Engineer
07
Agentic Systems
Built & shipped
- MCP
- Tool-calling agents
- Logic Apps
- Human-in-the-loop
↳ MCP Platform agent · 2x2x2 reporting agent with Teams review at Shell

ai-platform.stacklayer 01 / 07

Each layer depends on the reliability of the one beneath it.

06Engineering Projects

Concrete systems — designed, deployed and debugged.

An AI microservices platform an operator can interrogate through tool-calling agents, self-hosted LLM inference on Kubernetes, and an ML system that catches cloud cost spikes before the invoice does.

FLAGSHIP · 01mcp-platform-eks-gitops

MCP Platform

Cloud-Native AI Microservices on AWS EKS

GitOps + IRSA + MCP + LLM Tool Calling

View on GitHub

A 10-service AI microservices platform on AWS EKS, delivered through Argo CD with ordered sync waves, secured with IRSA, and operable through a Claude tool-calling agent that queries live platform services to answer operator questions.

services on EKS: 10services on EKS
ordered Argo CD sync waves: 7ordered Argo CD sync waves
static AWS credentials (IRSA): 0static AWS credentials (IRSA)
gateway replicas · HPA @ 70% CPU: 2–10gateway replicas · HPA @ 70% CPU

ns/mcp-platform

wn = Argo CD sync wave · traffic enters via /api/* on the gateway

Zero-downtime deployments for 10 services through a 7-wave ordered Argo CD sync
IRSA for mcp-control-plane — OIDC-federated STS token exchange, no static AWS credentials
LLM tool-calling agent (Claude Sonnet 4.6) answers operator questions by querying live platform services
API gateway autoscaling 2→10 replicas at 70% CPU · Kustomize overlays: EKS production, MicroK8s development

Terraform
AWS EKS
Argo CD
Kustomize
IRSA
OIDC
STS
HPA

FLAGSHIP · 02ai-inference-kubernetes-vllm

AI Inference on Kubernetes

vLLM · Gemma Model Serving

Self-hosted LLM · OpenAI-compatible API · SRE triage

A Kubernetes-hosted inference service for Google Gemma 3 1B on vLLM, exposing an OpenAI-compatible REST API so applications can consume self-hosted text generation — adapted from the upstream GPU deployment to run on ARM64 CPU within a 6 GiB memory budget.

model served by vLLM: Gemma 3 1Bmodel served by vLLM
compatible REST API: OpenAIcompatible REST API
container memory budget: 6 GiBcontainer memory budget
CPU execution · no cloud needed: ARM64CPU execution · no cloud needed

View on GitHub

triage-request.sh

illustrative

1# OpenAI-compatible endpoint served by vLLM in the cluster2curl -s http://vllm-gemma:8000/v1/chat/completions \3  -H "Content-Type: application/json" \4  -d '{5    "model": "gemma-3-1b",6    "messages": [{7      "role": "user",8      "content": "payment-api: CrashLoopBackOff, exit 137. Cause?"9    }]10  }'11 12# → diagnostic suggestion in choices[0].message.content

inference-pod.yaml

illustrative · simplified

1# what keeps the inference Pod recoverable2volumes:3  - name: model-cache        # weights survive Pod replacement4    persistentVolumeClaim:5      claimName: vllm-model-cache6readinessProbe:7  httpGet: { path: /health, port: 8000 }8resources:9  limits:10    memory: 6Gi              # ARM64 CPU execution budget

Kubernetes-hosted vLLM inference service with an OpenAI-compatible REST API
AI-assisted incident triage: simulated application failures in, diagnostic suggestions out
Persistent model storage reuses downloaded weights across Pod replacements
Health checks let Kubernetes manage recovery of the inference Pod

Kubernetes
vLLM
Gemma 3
OpenAI API
PVC
Health probes
ARM64

finops · cost-intelligence

synthetic data

30-day spend: $14,132
Forecast (30d): $13,166
Budget burn: 88%
Anomalies: 1

service: EC2 · daily cost observed expected band

ANOMALY DETECTEDday 47

Service: EC2 (example)
Expected: $430
Observed: $1,268
Potential impact: +$838/day

Service breakdown · 30d

EC258%
RDS17%
S313%
Lambda7%

FLAGSHIP · 03cloud-cost-spike-detector

Cloud Cost Spike Detector

FinOps ML · MLOps · Time-Series Anomaly Detection

An end-to-end ML platform that detects abnormal cloud cost spikes from billing data — enabling proactive FinOps monitoring, forecasting and budget governance, and estimating financial impact per service.

model: IsolationForest · 200 estimators · contamination 0.05
features: rolling mean / std · % change · deviation
serving: FastAPI · /detect · /detect/summary
delivery: Docker Compose · GitHub Actions CI (pytest)

Burn-rate tracking
Service-level visibility
Forecasting
Budget governance
CI/CD integration

View on GitHub

architecture · billing → decisionbatch scoring · REST serving

07System Design

System Design Lab. Watch the platform recompose.

Four reference architectures for the kinds of systems I build and operate. Switch modes to morph the topology, then select any component to see why it’s there.

01kubernetes-platform.arch

reference architecture

design notes

Spread replicas across zones; PodDisruptionBudgets protect capacity during node drains.
Rolling updates gated by readiness probes, so a bad build never takes traffic.
Keep state out of the cluster — managed databases with restore drills.

select any component for its role

signals to watch

p99 latency
error rate
saturation
pod restarts

Generic reference designs to show how I reason about systems — not diagrams of any employer's production environment.

07.b · interface

Prefer a terminal? Query the portfolio like a platform.

A scripted command engine that answers and navigates locally. It sits behind a small backend interface, so a grounded LLM can be plugged in later without touching the UI.

ASK THE PLATFORM

local · no external api

connected to paras@platform · local engine · type help

08Stack

Technical breadth, mapped as a dependency graph.

No skill bars. Each technology shows where it was actually used — a role, a project, or the résumé skills list — and what it connects to.

Languages

Python
Bash
PowerShell
Go
C++
Rust

IaC

Terraform
Packer
Ansible
Pulumi

Cloud

AWS
Azure
GCP
VPC / VNet
DNS & Routing

Containers

Kubernetes
EKS
AKS
GKE
Docker
Helm
Crossplane

Delivery

GitHub Actions
Azure DevOps
GitLab CI/CD
Argo CD
GitOps
Kustomize

Observability

Prometheus
Grafana
Dynatrace
Power Automate

Model Serving

vLLM
Gemma 3
OpenAI API
MLOps

AI & Agents

Azure AI Foundry
AWS Bedrock
MCP
Tool Calling
Logic Apps
RAG
LangChain
LLMOps

used in a role or project listed skill

08.b · credentials

Certifications

$ ls ~/certifications | wc -l → 7

CKA✓ certified
Certified Kubernetes Administrator
CNCF
HCTA0-003✓ certified
HashiCorp Certified: Terraform Associate
HashiCorp
AZ-400✓ certified
Designing and Implementing Microsoft DevOps Solutions
Microsoft
AZ-104✓ certified
Microsoft Certified: Azure Administrator
Microsoft
AI-102✓ certified
Microsoft Azure AI Engineer Associate
Microsoft
ML✓ certified
Machine Learning Specialization
DL✓ certified
Deep Learning Specialization
kubernetes · terraform · azure · devops · ai · ml

08.c · foundations

Education

Graduated 2023
Associate Degree in Computer Science and Programming
MIT (Massachusetts Institute of Technology)
USA
Graduated 2019
Applied Science (AMM)
Conestoga College
Canada
Graduated 2015
Bachelor of Engineering
RK University
India

09Connect

Let’s build systems that scale.

Interested in platform engineering, SRE, AI infrastructure, distributed systems and production GenAI platforms.

$pbhanderi.ai@gmail.com

LinkedIn GitHub Résumé

Bangalore, India · IST (UTC+5:30)

FLAGSHIP · 02ai-inference-kubernetes-vllm

AI Inference on Kubernetes

vLLM · Gemma Model Serving

Self-hosted LLM · OpenAI-compatible API · SRE triage

model served by vLLM: Gemma 3 1Bmodel served by vLLM
compatible REST API: OpenAIcompatible REST API
container memory budget: 6 GiBcontainer memory budget
CPU execution · no cloud needed: ARM64CPU execution · no cloud needed

View on GitHub

triage-request.sh

illustrative

1# OpenAI-compatible endpoint served by vLLM in the cluster2curl -s http://vllm-gemma:8000/v1/chat/completions \3  -H "Content-Type: application/json" \4  -d '{5    "model": "gemma-3-1b",6    "messages": [{7      "role": "user",8      "content": "payment-api: CrashLoopBackOff, exit 137. Cause?"9    }]10  }'11 12# → diagnostic suggestion in choices[0].message.content

inference-pod.yaml

illustrative · simplified

1# what keeps the inference Pod recoverable2volumes:3  - name: model-cache        # weights survive Pod replacement4    persistentVolumeClaim:5      claimName: vllm-model-cache6readinessProbe:7  httpGet: { path: /health, port: 8000 }8resources:9  limits:10    memory: 6Gi              # ARM64 CPU execution budget

Kubernetes-hosted vLLM inference service with an OpenAI-compatible REST API
AI-assisted incident triage: simulated application failures in, diagnostic suggestions out
Persistent model storage reuses downloaded weights across Pod replacements
Health checks let Kubernetes manage recovery of the inference Pod

Kubernetes
vLLM
Gemma 3
OpenAI API
PVC
Health probes
ARM64

Paras Bhanderi — Senior Site Reliability EngineerParasBhanderi

Automate the operational surface

Design for failure

Make infrastructure observable

Release frequency

Deployment failures

Manual deployment effort

Mean time to recovery

Production VMs remediated

Release cycle time

Operational toil

Microservices observed

Platform uptime

Shell

Gore Mutual Insurance

Kompusys Consultant Inc.

Apple Inc.

GitOps Platform Engineering

Cloud Infrastructure

Platform Engineering

Kubernetes

Distributed Systems

Model Serving

LLM Infrastructure

Agentic Systems

MCP Platform

AI Inference on Kubernetes

Cloud Cost Spike Detector

Prefer a terminal? Query the portfolio like a platform.

Certifications

Education

Let’s build systems that scale.

Automate the operational surface

Design for failure

Make infrastructure observable

Release frequency

Deployment failures

Manual deployment effort

Mean time to recovery

Production VMs remediated

Release cycle time

Operational toil

Microservices observed

Platform uptime

Shell

Gore Mutual Insurance

Kompusys Consultant Inc.

Apple Inc.

GitOps Platform Engineering

Cloud Infrastructure

Platform Engineering

Kubernetes

Distributed Systems

Model Serving

LLM Infrastructure

Agentic Systems

MCP Platform

AI Inference on Kubernetes

Cloud Cost Spike Detector

Prefer a terminal? Query the portfolio like a platform.

Certifications

Education

Let’s build systems that scale.

Paras Bhanderi — Senior Site Reliability Engineer