Skip to work
Rohann Keerthy / Portfolio

compiling shaders

A dark, quiet room and thirty thousand points of light.

000%

RKRK

Rohann Keerthy — Backend Software Engineer in Hyderabad, building agentic AI platforms, enterprise integrations and distributed backend systems.

Backend Software EngineerAgentic AI · Integrations · Distributed SystemsHyderabad · Telangana, India
Field · liveauto

Performance tier

I build the part nobody sees — agent platforms, integrations and the retrieval underneath.

Selected workStart a project
PythonJavaTypeScriptSQLFastAPISpring BootNestJSNode.jsLangChainRAGPostgreSQLMongoDBRedisCeleryDockerAWS S3PrometheusGrafana
PythonJavaTypeScriptSQLFastAPISpring BootNestJSNode.jsLangChainRAGPostgreSQLMongoDBRedisCeleryDockerAWS S3PrometheusGrafana
How I work001 / 006

Backends rarely fail on the happy path. They fail on the retry nobody wrote, the tenant who sees another tenant's row, the integration that loops forever at 3am. So I start at the edges: the read pattern, the isolation boundary, the failure mode. Then I instrument it, because a number you did not measure is a number you cannot improve. What ships is quiet, observed, and survives contact with fifty thousand users. 

50k+

Users served in production

across 1,000+ tenants

3

Production services owned

end-to-end

5

Enterprise integrations shipped

zero reported incidents

~2 yrs

Combined engineering experience

including internships

Selected work002 / 006

Systems in production, not prototypes.

Three production services at Cloudely and two shipped side projects. Each one has real users, a release history and numbers I can defend line by line.

Agent canvasdrag · test · publish
TRIGGERRESEARCHCRUDLLMPUBLISH
grounded · pii guarded5–10s
01In production

AI Agent Studio

Multi-tenant agent platform · Cloudely2025

An n8n-style drag-and-drop canvas for building agents, architected and shipped end to end. Tenant and user-scoped permissions, record CRUD tools, document parsing and insights, and Tavily-powered web research for lead research, generation and outreach. Prompt-injection and PII safeguards, grounded responses, and test/publish controls, with typical responses landing in five to ten seconds.

Role

Architecture and engineering, end to end

5–10s
typical agent response
Tenant
scoped permissions
PII
and prompt-injection safeguards
PythonFastAPIAgentic AILangChainLLM APIsTavilyMulti-tenancy
Connector platform5 integrations
PlatformGoogle WorkspaceSalesforceWhatsAppWooCommerceWeighbridge
oauth 2.0 · pkce · sha-256 loop guard0 incidents
02In production

Connector Platform

Enterprise integrations · Cloudely2025

Owned the connector platform and shipped five integrations: Google Workspace, Salesforce, WhatsApp Business, WooCommerce and Weighbridge. Salesforce runs bidirectional sync over OAuth 2.0/PKCE with conflict resolution and per-record SHA-256 loop prevention. WhatsApp carries real-time Sent/Delivered/Read webhooks plus in-app template creation. The Google Workspace connector lets tenant users reuse the identity they signed in with — Calendar read/write, bidirectional Sheets sync every thirty minutes, records created straight from Gemini meeting notes, Drive push notifications and pre-emptive token refresh.

Role

Platform owner

5
integrations shipped
Zero
reported production incidents
30 min
bidirectional sync cadence
OAuth 2.0PKCEWebhooksSalesforceMeta/WhatsApp GraphGoogle APIsHMAC-SHA256
Hybrid retrievalvector + bm25
queryembed
doc_48120.94
doc_09170.88
doc_22330.81
03In production

RAG Retrieval Chatbot

Hybrid retrieval backend · Cloudely2025

Owned the FastAPI backend and the embedding-model evaluation for a production RAG chatbot: LangChain, LLM APIs, MongoDB Atlas Vector Search and hybrid vector + BM25 retrieval. Tenant-aware Redis caching took roughly thirty percent off high-read API latency, Prometheus and Grafana halved mean time to detect, and Docker CI/CD carries regression coverage across sixty-plus pages. Presented the launch to over a hundred attendees.

Role

Backend owner, embedding evaluation

~30%
lower high-read API latency
~50%
lower mean time to detect
60+
pages of regression coverage
FastAPILangChainMongoDB Atlas VectorBM25RedisPrometheusGrafana
Daily brief20+ sources
Backend Engineer · career-site-0794
Python / FastAPI · career-site-1291
Platform Engineer · career-site-0387
04Open source

HireRadar

AI-powered job search automationPersonal

A FastAPI service that scrapes twenty-plus career sites, schedules searches, deduplicates postings and sends criteria-based alerts. Failures are isolated per source, so one broken career site never stops the full run. PostgreSQL persistence, URL and content-based deduplication, scheduled workflows, and pytest coverage over scraper and matching logic across wildly inconsistent DOM structures — a modular base for adding sources.

Role

Solo

20+
career sites scraped
Per-source
failure isolation
URL + text
deduplication
PythonFastAPISeleniumPostgreSQLpytest
RetailSparkFlutter · Play Store
9:41

Groceries,
hyperlocal

Produce
Dairy
Bakery
Pantry
Place order₹248
05Live on Google Play

RetailSpark

Hyperlocal grocery delivery appPersonal

A grocery-delivery app built and published to the Google Play Store for real-world use in Bangalore. Flutter/Dart client wired to backend APIs for product catalogue, cart, order placement, loading and error states, and order-status workflows.

Role

Client and API integration

Live
Google Play release
Bangalore
real-world usage
End-to-end
catalogue to order status
FlutterDartREST APIsGoogle Play

Want the architecture walkthrough?

Book a callRead the code
Specimens003 / 006

Modelled by script, not by hand.

Backend is the day job; this is the after-hours half. Three objects generated in Blender from a checked-in Python script, exported as glTF and lit in the browser. Drag one — the whole set is 222 KB over the wire.

Faceted shardloading · 13 KB
01 / 03
Technique

Convex hull of a perturbed Fibonacci lattice

Displacing a dense sphere and hoping for flat faces gives a lumpy rock. A convex hull is planar by definition — which is what a cut crystal actually is — so point count controls facet size directly.

Blender

Authored headlessly

bpy, no GUI

glTF

Shipped as GLB

position + normal only

0

Texture files

material written in code

222 KB

All three, gzipped

lazy-loaded

Approach004 / 006

Start at the failure path.

How I actually work on backend systems: constraint first, failure path second, observability before launch, and ownership after it.

01Before code

Find the real constraint

Latency, tenancy or trust — most backend work is one of the three wearing a disguise. I get to the read pattern, the isolation boundary and the failure mode before writing an endpoint.

1 / 4
02Design

Design for the failure path

Per-source isolation so one bad integration cannot stop a run. Idempotent retries, conflict resolution, loop prevention. The happy path is the easy half and almost never the expensive one.

2 / 4
03Delivery

Ship it observable

Nothing goes out without metrics and dashboards attached. Prometheus and Grafana halved mean time to detect on the RAG service; that number only exists because it was instrumented on day one.

3 / 4
04Ongoing

Own it in production

End-to-end ownership from architecture and vendor setup through to production support. Three production services, five integrations, and no reported connector-platform incidents.

4 / 4
Craft005 / 006

Agents, retrieval, integrations, scale.

Four areas I own end to end, from architecture and vendor setup through to production support.

Detail01 / 04

Agent platforms with tool access, grounded responses and guardrails that hold: prompt-injection and PII safeguards, scoped permissions, and test/publish controls before anything reaches a tenant.

Agentic AILangChainLLM APIsOpenAIGeminiTavily
Index006 / 006

On the record.

Nearly two years of combined engineering experience, internships included. I would rather show you a running system than a longer list.

Aug 2025 — Present·01

Software Engineer

Cloudely

Architected the AI Agent Studio, owned the connector platform and five integrations, owned the FastAPI backend for a production RAG chatbot, and built event-driven workflow and scheduling infrastructure on Celery/Redis.

Apr 2025 — Jul 2025·02

Software Engineer Intern

Cloudely

Shipped production Cashfree payment and WhatsApp Business integrations in Node.js/Express with webhooks for real-time transaction and delivery status. Restored email delivery by diagnosing misconfigured cron jobs, and improved PostgreSQL query performance by ~40% through indexing and pagination.

Aug 2024 — Jan 2025·03

Backend Developer Intern

Ezynest LLC

Cut REST API response time by ~30% building and optimising Java/Spring Boot services with Spring Data JPA and OOP/SOLID design. Improved PostgreSQL query performance by ~40% via normalised schemas and indexing, and integrated Jira and Okta SSO.

Nov 2021 — Jun 2025·04

B.Tech, Computer Science (AI & ML)

Mahatma Gandhi Institute of Technology, JNTU Hyderabad

CGPA 7.89/10. Coursework across data structures and algorithms, operating systems, software engineering, computer networks, artificial intelligence, machine learning and deep learning.

Toolchain
PythonJavaTypeScriptSQLFastAPISpring BootNestJSNode.jsLangChainRAGPostgreSQLMongoDBRedisCeleryDockerAWS S3PrometheusGrafana
Also on record

250+

DSA problems solved

LeetCode · HackerRank · Codeforces

600+

GitHub contributions

last twelve months

AWS

Code, Build and Deploy

Coursera certification

7.89

CGPA out of 10

B.Tech CSE (AI & ML)

Full history on LinkedInCode on GitHub