B.Tech CSE (AI & ML) @ VIT Bhopal · Building things people didn't ask for, then documenting them like they did

I'm a backend-leaning fullstack developer who has spent the last few years convincing computers to do things they were clearly not designed to do. I optimize things nobody profiled, automate things nobody asked to skip, and self-host things a $5 SaaS plan would've handled fine — because paying for infrastructure is for people who don't own a spare VPS and several regrets.

Currently studying AI & ML formally, while my actual day-to-day is queues, sandboxes, and yelling at Docker. Balance.

A Codeforces-style online judge, built from scratch because apparently "just use an existing judge" was too easy.

This is the one I'll actually defend in an interview. Full Turborepo monorepo, real distributed-systems problems, real 2 AM debugging sessions:

• Next.js frontend talking to an Express API, which writes submissions to Postgres (Prisma + Neon) and enqueues jobs onto Redis via BullMQ
• A worker process claims jobs, spins up Docker containers to compile and run untrusted user code against test cases — network isolated, resource-capped, no funny business
• Verdicts flow back through the pipeline: PENDING → RUNNING → ACCEPTED / WRONG_ANSWER / TLE / RUNTIME_ERROR / COMPILATION_ERROR — basically a tiny bureaucracy for your code's failures
• Deployed on an Oracle ARM64 VM via GitHub Actions + PM2, because I like my deployments the way I like my sleep schedule: automated, so I don't have to think about it

Also survived a personal gauntlet of monorepo hell: Prisma custom output paths that didn't resolve, Turbo caching env vars into oblivion, a missing dist/ folder, tsx ghosting me in production, and a JWT_SECRET that GitHub Actions simply forgot existed. Character-building, allegedly.

Still in progress, because real projects don't ship "done" — they ship "good enough to demo, terrifying enough to keep improving": per-testcase result storage, one-container-per-submission execution, an actual code editor instead of a glorified textarea, and security hardening beyond "please don't hack this, I'm one person."

Offline-First AI Voice Assistant — A privacy-first PWA that runs a WASM-powered Whisper model client-side, because sending someone's voice to a random API felt like a design flaw, not a feature. Web Workers keep the UI thread responsive while the model does the heavy lifting; Next.js API routes keep the LLM keys away from anyone with DevTools open.

Real-Time Collaborative Whiteboard — A custom WebSocket server broadcasting drawing events to a canvas rendering engine, because Figma clones are apparently a rite of passage now.

ShellShockedJS — A POSIX-compliant shell, written in Node.js, via the CodeCrafters "Build Your Own Shell" challenge. Custom parser, built-ins, process execution, the works. Yes, I wrote a shell in the language most known for npm installing half the internet. No, I don't see the irony.

Languages: JavaScript/TypeScript, Python, Java, C++

Backend & Data: Node.js, Express, FastAPI, PostgreSQL, MongoDB, MySQL, Redis, Prisma, BullMQ, WebSockets

Frontend: React, Next.js, Tailwind CSS

Infra & Tooling: Docker, AWS (EC2), GitHub Actions, Vercel, Linux, Git

AI-assisted dev: Claude Code, Gemini CLI — yes, I use AI tools to build things, and no, that's not a personality flaw, that's called using your tools

Four years of daily-driving Arch Linux, with a stint in Gentoo compiling my own kernel because apparently pre-built binaries were too mainstream for me.

Most recent flex: recovered a fully kernel-panicked Arch install — busted initramfs after an interrupted pacman transaction — via chroot, a mkinitcpio rebuild, and reinstalling DKMS nvidia modules from a live USB. Nothing was backed up. Everything was fine. This is not a recommendation, it's a confession.

• Working through DSA with a self-built, structured curriculum (skipped arrays and sorting — some fundamentals just aren't for me)
• Competitive programming on CodeChef, and ranked 3135 (Global) in TCS CodeVita 2025 — not top of the leaderboard, but comfortably ahead of everyone who didn't show up
• Certified by IBM in Python for Data Science and React, which mostly proves I can follow a syllabus when properly incentivized

If it's broken, I probably broke it myself first, on purpose, to see how it works.