Skip to content
Open to SDE roles · 15-day noticeDetails

I build
distributed systems, AI agents, real-time apps, Go services, Next.js frontends
that hold up in production.

I'm Saurav — a software engineer with 3 years shipping Python, Go and Node.js microservices, Kafka pipelines and production GenAI (RAG, LangGraph agents, MCP) at 99.9% uptime.

SDE at Devkraft · Delhi NCR, India · B.Tech NSUT

Doctor portalNext.js · SSR
Video consultsWebRTC · Go signaling
Partner webhooksREST · gRPC
GATEWAY00:42:33
AI agentsLangGraph · MCP
Event workersKafka · Celery
DataPostgres · Redis

simulated traffic · the shape of the platform I build and run at Devkraft

Event pipelines

500K messages an hour.
None processed twice.

Kafka and Celery workers that survive redelivery, partial failure and traffic spikes. Every handler is idempotent, every retry backs off, and backpressure keeps consumers from drowning.

Idempotency keysRetries + backoffBackpressureDead-letter queues
Orders APIFastAPI producer
PaymentsGo producer
Webhooksat-least-once
Schedulercron jobs
billing-workers
consumer group · 6 partitions
lag 0
evt_7f3a order.createdp0 · offset 18,420 14ms
evt_7f5f payment.capturedp3 · offset 18,427 9ms
evt_7f5f payment.capturedp0 · offset 18,434 duplicate · skippeddup
evt_7fa9 invoice.generatep3 · offset 18,441 retry 2/5 · okretry
18,204
processed
311
dupes skipped
57
retried

Partitioned for scale

Keyed partitions keep per-entity order while consumers scale out.

Safe under redelivery

At-least-once delivery, exactly-once effects via idempotent writes.

Fails gracefully

Exponential backoff, DLQs and alerts instead of silent data loss.

Fast reads

Indexing and N+1 fixes cut database latency by 65%.

Production GenAI

Agents that call tools,
and cite their sources.

On a real-time medical platform, LangGraph agents plan, call MCP tool servers and ground every answer in patient records through hybrid RAG — behind a gateway with circuit breakers, so a slow model never breaks a consultation.

doctor-portal / patient / chat
DrDr. Mehra · consultation #4821
Summarise this patient's last 3 visits and flag any drug interactions.
LangGraph runtrace · run_9c1e
  1. plannerplan → fetch visits, check meds412ms
  2. mcp · records.search3 visits · 12 documents
  3. retrieveBM25 + vector → 24 chunks → rerank top 5—
  4. mcp · drug_db.interactions2 medications checked—
  5. respondanswer + 2 citations—
● prompt cache hittokens 3.2kcircuit: closed

LangGraph orchestration

Stateful agent graphs with planning, tool use and retries.

MCP tool servers

Typed tools for records, drugs and scheduling — reusable by any agent.

Hybrid RAG

BM25 + vector search (pgvector) with reranking to cut hallucinations.

40% cheaper inference

Prompt optimisation and caching, tracked on Grafana cost dashboards.

Selected projects

Things I've shipped.
End to end.

GitHub
Devkraft · production

Medical AI Platform

Real-time diagnostic assistant: video consults with an AI agent working alongside the doctor.

  • Agentic platform behind an API gateway with circuit breakers and health checks
  • Patient-record RAG with hybrid search (BM25 + vector) and reranking
  • Go WebRTC signaling — doctor handoff under 200ms
  • Next.js doctor & patient portals on Django REST + FastAPI
LangGraphMCPRAGGoNext.js
Personal

Transcendify

Distributed video-transcoding queue that turns uploads into adaptive HLS streams.

  • Python workers on Celery + Redis handling 1,000+ videos/day
  • FFmpeg HLS ladders (1080p / 720p / 480p) served via S3 + CloudFront
  • Sub-500ms enqueue latency
PythonCeleryRedisFFmpegAWS
Self-hosted

This website

Next.js static export on my own Oracle Cloud ARM VM — no Vercel.

  • nginx + Let's Encrypt TLS, UFW + fail2ban hardened host
  • Docker Compose stacks, Authelia SSO, self-hosted Headscale VPN
  • Encrypted nightly restic backups with push alerts
Next.jsnginxLinuxDocker

Impact

Measured in production,
not in slides.

3 yrs

shipping production backends.

Microservices, event pipelines, real-time infra and GenAI — designed, deployed, on-call and improved.

_
99.9%

uptime on Kubernetes.

LiveRolling update Healthy

Rolling updates, HPA autoscaling and alerts that page before users notice — 70% faster recovery.

10K+

concurrent WebSockets.

+9K
−65%

database latency.

before · 120ms
after · 42ms
REST → gRPC
−40%

p99 latency.

Deadline propagation, zero-downtime rollout.

−40%

LLM inference cost.

500K+

messages an hour — none processed twice.

p0
p1
p2
p3
p4
p5

Observable by default.

Every service I ship has metrics, traces and dashboards before it has users.

PrometheusGrafanaOpenTelemetry

Stack

Tools I ship with.
Picked per problem.

  • Python
  • FastAPI
  • Django
  • Go
  • Node.js
  • NestJS
  • TypeScript
  • React
  • Next.js
  • GraphQL
  • WebSockets
  • PostgreSQL
  • Redis
  • MongoDB
  • Kafka
  • RabbitMQ
  • Celery
  • Docker
  • Kubernetes
  • Helm
  • GitHub Actions
  • Terraform
  • LangChain
  • LangGraph
  • Claude
  • Gemini
  • Hugging Face
  • Prometheus
  • Grafana
  • OpenTelemetry
  • nginx
  • FFmpeg

Experience

Where I've worked.
What I shipped there.

  1. Jun 2024 — PresentNOW
    Gurugram, India

    Devkraft

    Software Development Engineer

    Building and operating a 20+ service platform: Python, Go and Node.js microservices, Kafka pipelines, real-time infrastructure and production GenAI — on Kubernetes at 99.9% uptime.

    Show full notes
    • Python microservices (FastAPI, Django REST) and Go services (net/http, Gin) handling 50K+ requests/day across 20+ services.
    • Fault-tolerant Kafka + Celery/RQ pipelines at 500K+ messages/hour — backpressure, retries with backoff, idempotent handlers safe under redelivery.
    • Migrated inter-service calls from REST to gRPC with deadline propagation, cutting p99 latency ~40% with zero-downtime rollouts.
    • Real-time WebSocket infrastructure (Redis Pub/Sub) sustaining 10K+ concurrent connections.
    • Cut database latency 65% via PostgreSQL/MongoDB indexing, ORM N+1 fixes and schema work.
    • RAG pipelines (LangChain, pgvector) and LangGraph agents with MCP tool-calling; prompt optimisation + caching cut inference cost 40%.
    • HLS adaptive-bitrate streaming (FFmpeg, CloudFront) — 60% less buffering.
    • Next.js / React (TypeScript, SSR/SSG) apps for 50K+ daily users; Docker + Kubernetes with Prometheus, Grafana, OpenTelemetry — 70% faster incident recovery.
  2. Aug 2023 — Mar 2024
    Remote

    Alyssum Global Services

    Software Engineer Intern

    Shipped REST APIs on AWS and the CI/CD that deploys them.

    Show full notes
    • REST APIs (FastAPI, Node.js/Express, Go/Gin) on AWS Lambda and EC2 with JWT auth — 45% faster responses.
    • CI/CD with GitHub Actions + Docker cut deployment cycles 40%; Terraform + CloudWatch sped up incident response 35%.
  3. Aug 2019 — Jul 2023
    CGPA 8.3 / 10

    NSUT East Campus, Delhi

    B.Tech, Electronics & Communication

    Microsoft Ideathon '21 — Top 8 in India. HackNSUT — Top 10 of 300+. HackTheLeague (IIT) — Top 50 of 1000+.

For recruiters

Questions.

The things people usually ask before the first call.

Prefer to just talk?

sauravsharma3001@gmail.com

Email me
01What roles are you looking for?

Software Engineer / SDE-1 or SDE-2 roles in backend, full-stack or GenAI engineering — anywhere distributed systems and production reliability matter.

02What is your notice period?

15 days. I can start quickly.

03Where are you based?

Delhi NCR, India. Happy with on-site in Delhi/Gurugram, hybrid, or fully remote.

04What's your strongest stack?

Python (FastAPI, Django) and Node.js (NestJS) on PostgreSQL, Redis and Kafka, with Go for high-throughput services and React/Next.js on the front. For GenAI: LangChain, LangGraph, MCP and RAG with pgvector.

05Can I see code?

Yes — github.com/saurav470. Production work at Devkraft is private, but I'm happy to walk through architecture, trade-offs and incidents in an interview.

06How do I reach you fastest?

Email sauravsharma3001@gmail.com or message me on LinkedIn. I usually reply the same day.

Let's build something
that stays up.

Emailreplies same dayLinkedInsaurav-kumar-sharma-sks
OPEN TO WORK00:42:33
GitHub@saurav470RésuméPDF · updated Nov 2026
Email meDownload résumé

Delhi NCR, India · 15-day notice · open to remote