Kunal Jha

Backend & Distributed Systems Engineer.

Built systems at

Ultimate Kronos Group

Amazon

Accenture

Anthropic certified ×3

backend

Distributed-systems engineer who keeps p99s honest and regions independent.

see the systems →

{"role":"SE3","focus":["MCP","RAG","agents"]}
tools: get_pods · tail_logs · search_jql · soql_query
claude.messages.create({ model, tools, messages })
plan → search → rank → cite
human_in_loop = true
retry(idempotent=True, dlq="triage.dlq")
if severity >= P2: page(oncall)
@Tool("rollout_status") requireReadScope(ns)
embed(chunk) → pgvector → top_k=8
guard.allowlist(fields) ; audit.log(call)
p99 < 120ms · error_budget: 99.9%
kubectl get pods -n payments | grep -v Running
SELECT Id, Status FROM Case WHERE Priority = 'High'
✓ grounded · ✓ approved · ✓ shipped
{"role":"SE3","focus":["MCP","RAG","agents"]}
tools: get_pods · tail_logs · search_jql · soql_query
claude.messages.create({ model, tools, messages })
plan → search → rank → cite
human_in_loop = true
retry(idempotent=True, dlq="triage.dlq")
if severity >= P2: page(oncall)
@Tool("rollout_status") requireReadScope(ns)
embed(chunk) → pgvector → top_k=8
guard.allowlist(fields) ; audit.log(call)
p99 < 120ms · error_budget: 99.9%
kubectl get pods -n payments | grep -v Running
SELECT Id, Status FROM Case WHERE Priority = 'High'
✓ grounded · ✓ approved · ✓ shipped
{"role":"SE3","focus":["MCP","RAG","agents"]}
tools: get_pods · tail_logs · search_jql · soql_query
claude.messages.create({ model, tools, messages })
plan → search → rank → cite
human_in_loop = true
retry(idempotent=True, dlq="triage.dlq")
if severity >= P2: page(oncall)
@Tool("rollout_status") requireReadScope(ns)
embed(chunk) → pgvector → top_k=8
guard.allowlist(fields) ; audit.log(call)
p99 < 120ms · error_budget: 99.9%
kubectl get pods -n payments | grep -v Running
SELECT Id, Status FROM Case WHERE Priority = 'High'
✓ grounded · ✓ approved · ✓ shipped
{"role":"SE3","focus":["MCP","RAG","agents"]}
tools: get_pods · tail_logs · search_jql · soql_query
claude.messages.create({ model, tools, messages })
plan → search → rank → cite
human_in_loop = true
retry(idempotent=True, dlq="triage.dlq")
if severity >= P2: page(oncall)
@Tool("rollout_status") requireReadScope(ns)
embed(chunk) → pgvector → top_k=8
guard.allowlist(fields) ; audit.log(call)
p99 < 120ms · error_budget: 99.9%
kubectl get pods -n payments | grep -v Running
SELECT Id, Status FROM Case WHERE Priority = 'High'
✓ grounded · ✓ approved · ✓ shipped

<ai builder>

Wires LLMs into real infrastructure — MCP servers, agents, Claude in the ticket pipeline.

see the AI work →

@KafkaListener(topics = "epod.delivery.events", groupId = "verification")
public void onDelivery(DeliveryEvent event) {
    if (seen.putIfAbsent(event.id(), Boolean.TRUE) != null) return;      // idempotent by design
    retry.execute(ctx -> verifier.verify(event));                         // retries before features
    metrics.timer("epod.verify.latency").record(ctx.elapsed());           // measure, then optimise
}

@Tool(name = "get_pods", description = "List pods in a namespace (read-only)")
public List<PodSummary> getPods(@Param("namespace") String ns) {
    guard.requireReadScope(ns);
    return kube.pods().inNamespace(ns).list().stream().map(PodSummary::from).toList();
}

My Expertise

Backend
Development

Java, Spring Boot and REST — five years of production services at Ultimate Kronos Group, Amazon and Accenture. Versioned APIs, clean data models and code that survives its second team.

Distributed Systems
Multi-region, event-driven

SQS/SNS, Kafka, RabbitMQ and DynamoDB global tables. Idempotent consumers, retries, DLQs and honest dashboards — so a regional incident degrades gracefully instead of stopping deliveries.

AI Engineering
MCP, agents, RAG

MCP servers for Kubernetes, Jira and Salesforce, Claude inside the support workflow, an issue-context agent over Confluence and Jira, and Claude Code skills that build and hot-patch JARs into Kubernetes pods — with three Anthropic certifications behind it.

My sweet spot is the layer between “it works on one box” and “it works across regions.”

— Kunal Jha, on the part of the job he actually enjoys

// stack

The toolbelt

What I actually reach for in production — hover to pause.

Java17 · primary

Spring Bootservices · REST

Kafkaevent streaming

RabbitMQtopic listeners

PostgreSQLrelational

Rediscaching

DynamoDBglobal tables

Elasticsearchsearch · logs

AWSSQS · SNS · Lambda · ECS

Dockercontainers

Kubernetesorchestration

Java17 · primary

Spring Bootservices · REST

Kafkaevent streaming

RabbitMQtopic listeners

PostgreSQLrelational

Rediscaching

DynamoDBglobal tables

Elasticsearchsearch · logs

AWSSQS · SNS · Lambda · ECS

Dockercontainers

Kubernetesorchestration

GitHub ActionsCI/CD

Gitversion control

Linuxshell · sysadmin

IntelliJ IDEAdaily driver

Reactengineering tools

JavaScripttooling · UI

Pythonscripts · agents

JiraMCP server

SalesforceMCP server

Grafanadashboards

PostmanAPI testing

Claude · MCPAnthropic API · agents

GitHub ActionsCI/CD

Gitversion control

Linuxshell · sysadmin

IntelliJ IDEAdaily driver

Reactengineering tools

JavaScripttooling · UI

Pythonscripts · agents

JiraMCP server

SalesforceMCP server

Grafanadashboards

PostmanAPI testing

Claude · MCPAnthropic API · agents

First, we need a
problem…

We run a Java backend that has grown into forty microservices. Deploys are scary, and a regional outage last quarter took checkout down for an hour.

Alright — first I'd map the failure domains. What has to stay up if one region disappears? That decides where idempotency, retries and replicated state need to live.

How do you usually approach that?

The way I did at Amazon: region-local processing, async messaging with SQS/SNS, idempotent consumers with DLQs, and DynamoDB global tables for shared state. We measured roughly 30% higher reliability.

multi-region-logistics.svg

the shape of it — region-local processing, replicated state, retry + DLQ (from the case study)

And the scary deploys?

Split by domain, version the APIs, put contracts on the events. That migration cut latency 25–35% and manual verification 30–40%. Then wire in observability, so you trust dashboards instead of guesses.

We are also being asked to “add AI” somewhere.

Then add it where it saves your engineers time. I've built MCP servers for Kubernetes, Jira and Salesforce, and put Claude in the ticket pipeline with an engineer approving every reply. Grounded, permission-scoped — boring in the best way.

claude — kubernetes-mcp

› why is checkout slow in payments?▶ tool_use get_pods {"namespace":"payments"} 12 pods · 11 Running · 1 CrashLoopBackOff payments-worker-5c9…▶ tool_use tail_logs {"pod":"payments-worker-5c9…","lines":200} HikariPool-1 — connection is not available, request timed out after 30000ms✓ Likely cause: connection-pool exhaustion after the 10:40 deploy. Suggested: roll back payments-worker (mutating — needs your approval)

Claude debugging a namespace through the Kubernetes MCP server — read-only tools, a human approves anything that mutates

Perfect. When can we start?

Then, how I'd
work…

Reliability first

Idempotency, retries and DLQs before features.

Event-driven by default

Contracts over coupling; queues over cron.

Measure, then optimise

p99s and dashboards, not guesses.

Tooling is part of the craft

If it's manual twice, it becomes a script or an agent.

Now!

Things I write about

A few of the 25 posts I have published on LinkedIn — distributed systems, AWS, security and backend design. All 25 posts →

01 System Design

System Design Journey: Building a Concurrent System

Walking through the design of a concurrent system — the trade-offs, the failure modes, and what I'd do differently.

Jun 2026read on linkedin ↗ 02 Data & Messaging

Most Engineers Say Redis Uses LRU Eviction — It Doesn't, Exactly

Redis' eviction is approximated LRU with sampling, and that detail matters when you size a cache.

May 2026read on linkedin ↗ 03 AI

RAG & LLMs: What Actually Works in Production

Chunking, retrieval quality and grounding — the parts of a RAG pipeline that decide whether the answer is trustworthy.

May 2026read on linkedin ↗ 04 AWS

DynamoDB for Backend Engineers

Partition keys, access patterns and the mistakes that make single-table design hurt later.

Nov 2025read on linkedin ↗ 05 Data & Messaging

Kafka vs Amazon Kinesis

Ordering, retention, consumer models and cost — how to pick between the two for event streaming on AWS.

Oct 2025read on linkedin ↗ 06 Backend & APIs

API Design: Pagination Done Right

Offset vs cursor pagination, why offsets break at scale, and how to design stable cursors.

Sep 2025read on linkedin ↗ 07 Security

API Security Lessons from Amazon

Auth boundaries, least privilege and the checks that keep internal APIs safe at scale.

Sep 2025read on linkedin ↗ 08 Data & Messaging

Google MapReduce — The Paper That Started Big Data

A plain-English walkthrough of the MapReduce paper and why its ideas still shape data systems.

Sep 2025read on linkedin ↗ Blog

17 more posts

System design, backend, AWS, data, security and AI — filter by topic on the blog page.

open the blog →

Get In Touch

Let's build something resilient.

Hiring for a backend, distributed systems or AI engineering role? Need someone who's run multi-region services in production? Or just want to trade notes on Kafka, AWS and keeping p99s honest? My inbox is open.

> email ./linkedin ./call ./resume

@ Delhi, India$ kjha38017@gmail.com$ +91 9643221081