Sujit Kumar Maiti Backend & AI infrastructure

I build the machinery behind products.

Backend and AI-infrastructure engineer at Cravin Technologies, Dubai. I design the one constraint that keeps a system correct under retries and load: an idempotency key, a queue boundary.

Résumé (PDF) Contact me
Experience
4 years
Focus
Backend + AI infrastructure
Based in
Goa, India
Works
Remote, or on-site in India
Open to
Full-time roles

Core stack TypeScript · Node.js · NestJS · Python · FastAPI · PostgreSQL · Redis · AWS · Docker · LLM gateways

A sheet of paper. Flat, every request that lands on it runs a job, duplicates included. Creased, the sheet becomes a rigid shell: the first request with each key runs once and repeats roll off the crease.
0requests0jobs run0duplicates absorbed
A demo of how I stop duplicate jobs when requests are retried. Drag the crease, then replay.

Three systems, three creases.Each began as a flat sheet: a failure mode that would happen under load. The crease is the decision that made it hold.

I

Idempotent Task Grid

Python · FastAPI · Redis · Docker Compose · Asyncio
Problem

In distributed systems, client network timeouts or uncoordinated retries trigger duplicate execution of asynchronous background jobs, corrupting state and doubling compute cost.

The crease

Built an ingestion service using FastAPI and Redis. Incoming requests pass an Idempotency-Key header verified atomically via Redis SET NX before dispatching to asyncio.BackgroundTasks. Sliding-window rate limiting prevents request bursts.

Result

Prevents duplicate worker execution on network retries with sub-50ms ingestion latency.

Source on GitHub

clientFastAPIRedis SET NXasyncio worker
Plan · services
time →claim → runretry → stored job
Section · first request and its retry
II

LLM Observability Gateway

NestJS · React · Prisma · PostgreSQL · BullMQ · Redis · Prometheus · Grafana
Problem

Capturing granular inference telemetry (token counts, TTFT, model latency, error rates) in streaming LLM requests normally introduces database I/O latency and network jitter into the client stream.

The crease

Multi-provider routing gateway in NestJS supporting OpenAI-compatible and Anthropic endpoints over Server-Sent Events (SSE). Streams tokens immediately to the client while offloading inference telemetry payloads to Bull queues backed by Redis.

Result

Maintained low time-to-first-token while capturing structured audit logs and Prometheus metrics asynchronously.

Source on GitHub

clientgatewayOpenAI APIAnthropicBull → metrics
Plan · services
time →tokens to client (SSE)telemetry, queued
Section · one request (schematic)
III

Food Delivery Infrastructure

NestJS · TypeScript · PostgreSQL · Redis · Socket.IO · Turf.js · Stripe · Next.js
Problem

Multi-branch restaurant ordering requires determining exact branch delivery zone eligibility from coordinates, while real-time kitchen alerts and Stripe webhook retries must never cause duplicate orders.

The crease

NestJS microservices utilizing @turf/boolean-point-in-polygon for delivery boundary validation, partitioned Socket.IO rooms for kitchen dispatch, and cryptographic Stripe webhook signature verification with idempotent event handling.

Result

Reduced p95 response time by 25% and prevented duplicate webhook processing across branch workflows.

In production at Cravin

branch zonekitchen socketStripe, idempotent
Plan · zone, kitchen, payments

How I decide where the crease goes.

01

Idempotency over optimism

Networks fail, clients timeout, and users double-click. If a write endpoint is not designed for safe retries, it is broken by default.

02

Decouple persistence from client latency

Synchronous database transactions during live streams or webhook ingest create backpressure. Push telemetry and secondary workflows to asynchronous queues.

03

Quiet, verifiable architectures

Systems become reliable through clear boundaries and observability, not through overly clever abstractions.

Where the work happened.

Dec 2023 to Present

Full Stack Developer

Cravin Technologies LLC FZCO, Dubai, UAE
  • Architected backend microservices in NestJS & PostgreSQL for food delivery and booking products, cutting p95 response time by 25%.
  • Collaborated on merchant management interfaces in Next.js with live Socket.IO feeds and TanStack Query state caching.
  • Built delivery zone boundary calculations using Turf.js geospatial polygon validation and idempotent Stripe webhook processing.
  • Automated deployment pipelines and container orchestration with Docker and AWS.
Sep 2022 to Nov 2023

Associate Software Engineer

Coditas Solutions LLP, Pune, India
  • Engineered enterprise backend microservices in Node.js & PostgreSQL for high-throughput client platforms.
  • Co-architected a serverless background verification system on AWS Lambda with asynchronous event queues, Redis database caching, and S3 encryption.
  • Optimized complex PostgreSQL queries and index usage, reducing database CPU load during peak reporting spikes.
Aug 2022

Bachelor of Engineering, Computer Science

Goa University, Goa, India
Materials
Backend
TypeScript, Python, Node.js, NestJS, FastAPI, Express.js
Data
PostgreSQL, Redis (SET NX, Pipelines), MongoDB, BullMQ, Prisma
Cloud
AWS (Lambda, SES, S3, EC2), Docker, Prometheus, Grafana, CloudWatch
Interfaces
Next.js, React, TanStack Query, TanStack Table, Tailwind CSS
hello@sujitmaiti.cv

Open to full-time backend and AI-infrastructure roles: remote, or on-site in India.

Hiring? Send the role, team and location. I reply within 2 working days.

Write an email Book a call

GitHub LinkedIn Résumé (PDF) resume.json

Or send a message here