Neerav Jha.

Software Engineer at Yellow.ai building LLM-orchestrated conversational AI at enterprise scale — RAG pipelines, function-calling orchestration, and resilient APIs. I find production bottlenecks and kill them: 3.8s → 473ms on my last one.

✉  Get in touch ↓  Resume ⌥  github/neerav34
Noida, India open to SWE / SDE roles ships to npm & production
neerav — zsh — things I actually shipped
0ms
cache-hit p50 after perf fix
0×
throughput gain · 1.9 → 10.3 req/s
0+
LeetCode problems solved
0+
npm downloads · env-doctor
01

Where I've worked

Associate Software Engineer Yellow.ai SEP 2024 — PRESENT · BANGALORE
PROMOTED FROM INTERN · MAR 2025
  • Built and productionized multi-channel conversational AI (chat & voice) using RAG pipelines and LLM orchestration — function calling, tool routing, caching — improving response latency and containment rate.
  • Designed and shipped the platform's first ServiceNow integration (OAuth 2.0, event-driven sync, resilient REST APIs), improving API reliability and throughput while reducing mean time to recovery.
  • Added performance instrumentation across orchestration services to surface latency and reliability regressions before they hit customers.
Python Developer Intern Munsow Technologies JUN 2024 — NOV 2024 · REMOTE
  • Built backend services (Flask, MongoDB) for an AI-powered flashcard learning platform, integrating the OpenAI API for automated content generation.
  • Implemented RBAC and RESTful APIs supporting 1,000+ registered users and 150+ concurrent sessions.
Data & Machine Learning Intern NIELIT JUN 2023 — AUG 2023 · HARIDWAR
  • Developed classification models in Python achieving ~80% prediction accuracy; optimized MySQL/Oracle queries on 1TB+ datasets, improving execution efficiency by ~40%.
02

Things I've shipped

Scan it yourself · Live in production

QRDrop

Next.js · WebRTC · Socket.io · Service Worker (PWA)

Send a file between any two devices by scanning a QR code — it streams straight browser-to-browser over an encrypted WebRTC data channel. No upload, no cloud storage, no account. Works Android → iPhone → laptop, the exact gap AirDrop and Quick Share leave open.

  • Survives real interruptions — screen-wake lock, resume tokens, and byte-exact reconnection: killing a live link mid-file and re-verifying by sha256 is a CI test, not a hope.
  • Optional 6-digit PIN gates the file manifest itself, not just the bytes — a locked join leaks nothing, and a hard attempt limit (not the hashing) is the real defense.
  • 152 checks across 8 test suites — protocol, PIN, TURN, PWA, and real two-Chrome-tab transfers — plus a TURN relay fallback for networks that block direct P2P.
GitHub ↗ Live — scan with your phone ↗
※ free-tier hosting sleeps when idle — first request may take ~30s to wake (page pings it the moment you open it)
$ npm run test:e2e ✔ 3MB transfer · sha256 verified ✔ 64MB · link cut mid-flight · resumed ✔ 4 files, 1 session · each sha256 verified ✔ PIN-gated transfer · verified ────────────────────────────── ✔ 152 checks · 8 suites · 0 flaky
Featured · Live in production

LLM Gateway

Python · FastAPI · Redis · Vector DB · SQLite · SSE

An OpenAI-compatible multi-provider routing gateway with automatic failover, two-layer caching, rate limiting, and cost tracking — deployed and load-tested on real infrastructure.

  • Semantic caching — embeddings + cosine similarity with a threshold calibrated from measured data (paraphrases ~0.89–0.92 vs unrelated ~0.56), serving cached answers for reworded prompts.
  • Diagnosed a connection-pooling bottleneck (per-request TLS handshakes) — p50 latency 3.8s → 473ms, sustained throughput 1.9 → 10.3 req/s.
  • Sliding-window rate limiting and streaming with pre-first-byte provider fallback.
GitHub ↗ Live API ↗ Demo video ↗ Writeup on Medium ↗
※ free-tier hosting sleeps when idle — first request may take ~30s to wake
p50 · before3,802 ms
p50 · after pooling fix473 ms
−87% latency · one commit
$ python scripts/loadtest.py ────────────────────────────── ✔ p50 473ms · p95 520ms · p99 925ms ✔ throughput 10.3 req/s @ c=10 → X-Cache: SEMANTIC-HIT (sim 0.93)
Distributed systems · Live in production

taskq

Python · FastAPI · Redis · Distributed Workers

A distributed task queue — the same problem category as Celery, SQS, and BullMQ. Any app enqueues background jobs over HTTP; a pool of workers processes them reliably, even when a worker is killed mid-job.

  • Dead-worker rescue — job leasing with Redis TTL heartbeats: kill -9 a worker mid-job and another completes it within seconds; atomic claims prevent double-rescue.
  • Real failure handling — idempotent enqueue (SET NX), exponential-backoff retries into a dead-letter queue, priority scheduling, and backpressure (503 + Retry-After at queue-depth caps).
  • Load-tested at ~520 req/s enqueue, p50 28 ms locally — with honestly documented at-least-once semantics and known gaps, like a real system.
GitHub ↗ Live API ↗ Live dashboard ↗
※ free-tier hosting sleeps when idle — first request may take ~30s to wake
$ kill -9 $(pgrep -f worker-1) # mid-job lease:job_8f2 expired · no heartbeat → orphan detected · re-queued ✔ rescued by worker-2 · completed (3.1s) ────────────────────────────── $ python -m loadtest.enqueue_load ✔ 520 req/s · p50 28ms · p99 202ms ✔ 500/500 accepted · 0 jobs lost
Open source · 1,500+ downloads on npm

env-doctor

Node.js · CLI · GitHub Action · Vercel

"The ESLint of environment variables" — a CLI that catches missing and undocumented env vars before they hit production. Published on npm as @neerav34/env-doctor, currently at v1.0.10.

  • A full product surface, not just a script: npm package, GitHub Action for CI, VS Code extension, and a landing page.
  • Monorepo support, destructuring detection (const { VAR } = process.env), commented-code false-positive fixes — shipped across 10 releases.
npm ↗ GitHub ↗ Landing page ↗
$ npx @neerav34/env-doctor Scanning 214 files… ✔ DATABASE_URL .env.example ✓ ✔ REDIS_URL .env.example ✓ ✖ STRIPE_KEY missing from .env.example ✖ GROQ_API_KEY used but undocumented 2 problems found — fix before deploy
03

What I work with

◆RAG Pipelines◆LLM Orchestration◆FastAPI◆Vector Databases◆System Design◆Redis◆OAuth 2.0◆Async Python◆Node.js◆Docker

Languages

PythonJavaScriptJavaC++CSQL

Backend

FastAPIFlaskNode.jsExpress.jsREST API DesignOAuth 2.0

AI / LLM & Systems

RAG PipelinesLLM OrchestrationVector DBsSystem DesignAsync / Concurrency

Data & Infra

MongoDBPostgreSQLMySQLRedisSQLiteDockerAWSGit
04

Proof of work

Numbers that don't lie

1,100+ LeetCode problems solved — plus 5★ Python rating on HackerRank and 1,000+ problems across platforms.
−87% Production latency cut — diagnosed per-request TLS handshakes, fixed with a shared connection pool, verified under load.
1,500+ npm downloads of env-doctor across 10 releases — with CI, a GitHub Action, and a landing page kept in sync on every release.

Education

B.Tech — Computer Science & Engineering
HNB Garhwal University (A Central University)
SRINAGAR, UTTARAKHAND · GRADUATED AUG 2024
CGPA 8.0 / 10

How I work

I measure before I claim. Every performance number on this page came from a load test I wrote, and every project links to something you can run right now.

05 — LET'S BUILD SOMETHING

neeravjha444@gmail.com

Open to Software Engineer / SDE roles — India or remote. If you got this far, the fastest way to know if I'm your engineer is a 30-minute call.

GitHub LinkedIn Medium Resume ↓