Ramanujam Vaduganambi

Backend Engineer @Oraczen — LLM infra, observability, 100k-row pipelines

Hyderbad, India

deepakrama321@gmail.com

About

Most of my good work looks like deletions. A crash loop that wasn't a memory leak. A 15-minute pipeline that was waiting on the wrong model. A permission check is doing far too much thinking.

Skills

Python

FastAPI

Pydantic

Celery

Alembic

pytest

PostgreSQL

ClickHouse

Redis

RabbitMQ

Docker

Kubernetes

Helm

ArgoCD

GitHub Actions

Keycloak

OIDC

Casbin

JWT

LangChain

Langfuse

OpenTelemetry

MCP

Vector Search

Prompt Engineering

TypeScript

React

Next.js

Grafana

Sentry

Distributed Systems

Performance Optimization

Observability

Experience

Backend Engineer (AI Platform)

Oraczen

Dec 2024 to Present

Backend and AI-platform engineering across shared services: authorization/RBAC, LLM routing, observability, configuration, and a data-enrichment pipeline running 100k+ rows per run.
Rewrote Casbin permission resolution into an indexed lookup path — 229x–965x faster with byte-identical output, shipped across 17 consumer services.
Cut a production enrichment step from 15m14s to 35s and improved row completion from 9/10 to 10/10 by identifying model output-headroom limits and rerouting the workload.
Reduced enrichment-worker boot memory from 709MB to 427MB (40%) by removing eager heavyweight imports.
Diagnosed a platform-gateway crash loop as CPU throttling and event-loop blocking rather than memory exhaustion; traced repeated probe failures to a 150m CPU limit.
Migrated shared observability to Langfuse v3 and OpenTelemetry, centralising LLM-call telemetry behind a single facade.
Drove per-service database separation with independent Alembic migration chains.
Shipped security fixes across budget-enforcement bypasses, caller-JWT forwarding, static bearer-token fallbacks, and cross-tenant authorization assertions.

React JS Developer (Intern / Project Trainee)

Thought Process Systems LLP

Jun 2024 to Nov 2024