Software Engineer

Vishal Pundhir · India

I turn business problems into working systems.

And then I make sure they keep working.

Event-driven backends, queues and pipelines — the parts that decide whether a product survives its own traffic.

I move fast. I also care about getting it right.

Vishal Pundhir
Vishal PundhirBITS Pilani, 2024

2.3 yrs

building and running systems in production

1000+

videos a month through the pipeline I own

Writing code was never the job.

I use AI to multiply what I can do.

Engineering was always about solving the problem with the best tools your domain hands you. One of those tools now writes code. So I use it every day — to read more logs than I could, to draft three designs instead of defending my first idea, to try every path through a piece of test data. What it doesn't do is decide which number matters, or notice that the number everyone quotes has no source, or own it at 2am. That part is still the job.

BITS Pilani campus at dusk, the clock tower lit against a violet sky

Education

BITS Pilani

Bachelor of EngineeringClass of 2024

Pilani campus. Years of being surrounded by people who were better than me at something, which turned out to be the useful part.

About

I build AI products end to end — pipeline, backend, frontend, and keeping them running once they are live. Outside work I build tech for the everyday problems I notice around me. Learning by doing is what keeps making me better at the part that matters: working out what the real problem is.

Off the clock

Guitar and singing. Filmmaking. Football — playing, not watching.

Skills

What I work with.

Grouped by what I actually reach for. The three marked core are where most of my time goes.

Backend & APIs

Core

Where most of my work lives.

  • Node.js
  • Python
  • Express
  • NestJS
  • FastAPI
  • REST
  • MongoDB
  • MySQL

Queues & event-driven

Core

Kafka, SQS, FIFO ordering, idempotency, retries, dead-letter queues.

  • Kafka
  • SQS / FIFO
  • SNS
  • CDC events
  • Round-robin dispatch
  • Idempotent consumers

AI in production

Core

Shipped in real workflows, not demos — and the same tools run my own day. I measure whether they actually earn their cost.

  • Structured outputs
  • LangGraph
  • Agent loops & critics
  • RAG
  • Tool calling
  • MCP
  • LLM-as-judge
  • Evals & measurement
  • Speech in / out
  • Prompt versioning

Cloud & infra

Deploying it, running it, and being the one paged for it.

  • AWS
  • Lambda
  • Step Functions
  • ECS
  • EC2
  • S3
  • CloudFront
  • Nginx
  • PM2

Frontend

Enough to own a product end to end.

  • React
  • Next.js
  • TypeScript
  • Tailwind
  • Three.js / WebGL
  • React Native

Monitoring

The one I care about most. A system you can't see into is a system you're guessing about — and every number on this page came from having built the thing that reports it.

  • Grafana
  • CloudWatch
  • ClickHouse
  • Metabase
  • Loki
  • Structured logging
  • Alerting

Design

The creative side I never dropped.

  • After Effects
  • AE scripting
  • Premiere Pro
  • Illustrator
  • Figma

The list is the easy part. Picking one, and knowing why not the other, is the hard part — which is what the next section is actually about.

How I work

Same four steps, every time.

In this order. The tools change constantly; the sequence hasn't.

  1. 01

    The problem, and what the customer actually needs

    A ticket tells you what someone typed. I go find out what they were trying to do — which is often a different problem with a different answer.

  2. 02

    Break it into parts that can each be wrong

    If a failure can only be described as “it broke”, it was never broken down. Each piece should be able to fail on its own, visibly.

  3. 03

    Find the edge cases before a customer does

    The empty field. The message that arrives twice. The one dealer who uploads a hundred cars at once. These decide whether it survives.

  4. 04

    Design the system — then prove the number moved

    Pick the architecture, write down what it was chosen over, ship it, and measure it. Every number I claim comes with how I measured it.

Recent work

Problems I worked on.

Four from work, one from my own time. Open any one: what was asked, what it turned out to be, how I built it, and what changed. Written for someone who has never seen these systems — no jargon, and where a word is unavoidable I explain it.

Projects

Built on my own time.

Nobody asked for any of these. Each one started with something I wanted to understand, so the interesting part is usually one decision rather than the whole thing — that's the line marked ◆ on each card.

The ItineraryAI landing page — “Where group trip plans finally leave the group chat” over a night sky, with a Plan Your Trip button.
Live2026

ItineraryAI

Agentic trip planner

Say where you're going and for how long, and you get a day-by-day plan back. An agent pulls what it needs as it needs it — places, stays, current prices — and then a separate critic reads the finished plan with fresh eyes before you ever see it.

The critic loop stops after three attempts. Without a ceiling, a critic that keeps rejecting and a planner that keeps rewriting will happily burn tokens all night.

  • LangGraph
  • OpenAI
  • FastAPI
  • ChromaDB
  • Next.js
  • Express
  • MongoDB
  • Railway
The mini-claude web UI — a prompt to write and run a primes script, followed by the agent's reasoning, three tool calls, the answer, and the token cost for the turn.
Live2026

mini-claude

A coding agent, built from scratch

A small version of the AI coding assistants everyone now has open all day, built to find out how they actually work. Give it a task and it reads files, writes them, and runs commands until it's done. The page shows every step live — what it's weighing up, which tool it reached for, what came back, and what the turn cost.

It's public, so the shell it runs commands in works off a list of what's allowed rather than a list of what's banned. A list of dangerous commands is one you are permanently one entry short on.

  • Python
  • FastAPI
  • SSE
  • OpenAI Responses API
  • Docker
  • Railway

Smaller experiments — a 3D haunted house, a shirt you can spin and redesign, a Kafka implementation, a realtime drawing board — live on GitHub without a write-up. All of them are there.

Contact

Say hello.

Happy to talk about engineering, business problems, music, or football. If you're hiring: I want to own something real, with users and a number attached to it.

I'll happily argue the other side of anything on this page. Most of what I know came from being wrong out loud.