four years in production
Vol. 01 · 2022 — 2026
this notebook belongs to
bengaluru · karnataka · in mail · vijayrauniyar1818@gmail.com web · vijaygupta18.github.io git · github.com/vijaygupta18
If this notebook has reached you by accident, please drop me a note. There is some reasonably personal reading inside.
— VG
I build backends — the kind that can wake you up at 3 AM.
I have been doing this for four years, and I am currently
SDE 2 at Namma Yatri, part of a small team that grew the platform's
daily load sixty times over in two years while keeping the pager
mostly quiet. My title says SDE 2
, but in practice the work is part
feature engineering, part distributed systems, part infrastructure, and a
lot of staring at Grafana graphs at 2 in the morning.
I grew up writing Python, discovered Haskell at Juspay in 2023, and
today most of Namma Yatri is written in it. I prefer languages that make
it hard to do the wrong thing — Haskell, Rust, and the good parts of
TypeScript — and I enjoy the parts of the stack that most people call
boring
: Redis, Kafka, Postgres, and the zone-aware plumbing
underneath them.
I care most about three things, in this order: cost, latency, and the on-call page not firing. The rest of this notebook is essentially a record of that work — features I have shipped, systems I have helped build, and infrastructure I have helped tame. There are also a few side projects that kept me sane between incidents.
— V.
A short list, in no particular order, and with honest caveats. Every number
here is measured off real production graphs, not back-of-envelope estimates.
Where a number is approximate, I have said so. An asterisk means this
was a team win in which I played a meaningful part
— not I did
this alone
.
March 2024. Namma Yatri spun out of Juspay and I stayed on with the team. On day one we were doing roughly five thousand rides a day across Bengaluru, and the backend was a small, stubborn service written in Haskell. Today, as SDE 2, I help run a platform that serves three hundred thousand+ rides a day across multiple cities on the same Haskell core — but almost nothing else about it is the same.
In plain English, my job has been to keep the platform from falling over as it grows. The work falls into three parts.
I designed the multimodal journey planner — the feature that
lets a user type Majestic to Whitefield
and receive a stitched
itinerary across autos, metro, and bus with honest ETAs and
traffic-aware routing. It now ships in three cities. I also built large
parts of the real-time dynamic pricing engine, which decides what
an auto ride should cost at any given moment, based on demand, weather,
and the live conditions on a corridor.
The Redis / Valkey KV framework is the piece I am proudest of. It takes our real-time state — drivers, rides, prices — and stores it in a Redis-shaped structure that mirrors our Postgres schema. We run it at 5 million+ transactions a day and 5,000+ events per second, using table-level sharding, asynchronous operations, pipelining, and autoscaling that does not make you nervous. On top of that, I wrote an in-memory GTFS service with GraphQL preloading that reduced hot-path read latency by 60% at approximately 5,000 requests per second.
The platform now runs active-active on AWS and GCP, with cross-cloud Redis and Kafka routing, client-aware dispatch, and zero-downtime cross-cloud deploys. I led the Redis → Valkey migration (memory down 50%, cost down 40%), the Kyverno pod-zone injection that finally fixed our cross-AZ bill, and the cost work that quietly saved $126K a year on AWS. I own on-call for the core, and we have held 99.9% availability while the platform grew sixty times over.
Vishwakarma (p. 70) is an autonomous SRE agent. It runs sixteen parallel investigations for every alert and delivers an RCA PDF into Slack within a few minutes — cutting investigation time by 70%. ART (Automated Regression Testing) replays real production traffic against candidate builds, so we catch API regressions before a deploy, not after.
April 2023. I joined Juspay as an SDE on what would later become Namma Yatri. The stack was Haskell and a lot of Redis. This was where I got serious about types, and about treating the database as a system to be designed rather than a service to be called.
Two headline projects from that year:
We were paying too much for Postgres reads on data that did not really need to be strongly consistent. KV is a framework that keeps real-time rows in Redis, in a structure that mirrors the underlying table, and keeps them in sync through a drainer service. It reduced our database spend by 40% while making reads noticeably faster. It later scaled, at Namma Yatri, to 5 million+ daily transactions.
Originally, the drainer lived inside the application — which meant that a deploy of the app was also a deploy of the drainer, and a mistake in one could affect the other. I split the drainer out into its own service. Data now flows cleanly from Redis Streams into Postgres, and directly from Kafka into ClickHouse for analytics. With data being shaped at the edge rather than inside the app, deploys became far less risky.
June 2022. My first full-time role. My slice of the product was the backend for an AI-driven WhatsApp bot that matched blue-collar workers to jobs at Zomato, Swiggy, Zepto, and Uber. It was the kind of workload where every ten milliseconds you shave matters, because there is a real person on the other side waiting for a reply.
What I shipped:
It was a good year. I learnt how a real product behaves under a real load, and how different that is from a classroom assignment, where no one is waiting for your response.
Work is one thing. The best way I know to actually learn something is to build a small, slightly embarrassing version of it over a weekend. Pinned below are the projects I am happy to show someone.
An autonomous SRE agent. When you receive a page, it starts investigating, fanning out across Prometheus, Kubernetes, Elasticsearch, and the databases, and posts a full RCA PDF into Slack. Investigation time down 70%.
github ↗A drag-and-drop system-design practice environment. 30 components with real benchmarks (load balancer: 1M QPS, cache: 100K QPS), 35 design problems, and a traffic simulator with bottleneck detection.
github ↗A voice and video interview flow using Deepgram for STT and TTS, with real-time proctoring and automated scoring. Multi-tenant, and works with any OpenAI-compatible API.
github ↗A single console that queries Postgres and Redis across AWS and GCP at the same time. Parallel execution, side-by-side diff, three-tier RBAC, a Monaco editor, and password-gated destructive operations.
github ↗A Slack-native incident platform. Mention @argus and it triages the issue, assigns on-call by availability, and streams an LLM-powered RCA into a dashboard. Includes RBAC, audit trail, and reminders for stale tickets.
github ↗A CLI that lets you query any codebase using local LLMs (Ollama) with RAG. Ask a repository anything — nothing leaves your machine. Especially useful when you are new to a monorepo and would rather not interrupt a teammate.
github ↗A real-time Kubernetes cluster monitor covering pod health, resource usage, and an events viewer. Built with FastAPI and React, and deploys on a single pod.
github ↗An open-source project under @nammayatri. It crowd-sources bus-stop and route data using real-time GPS, and provides an admin API for corrections.
github ↗Sketched first as a mind-map, then listed in full below, because the structured list is more honest. ★ means deepest experience. · means I have shipped with it and would happily pick it up again. ~ means I am familiar with it around the edges.
Also on the bench: VS Code, GitHub Actions, Slack bots, Mermaid
diagrams, and more I will read the docs this weekend
than I care
to admit.
These are recommendations from former managers and teammates, originally posted on LinkedIn. I have pasted them here as they were written. The handwriting is mine; the words are not.
Vijay was a fantastic person to work with — multi-skilled, insightful, and with very strong problem-solving skills. His focus keeps everything moving smoothly, deadlines are met, and whatever project he is working on meets the highest standards. An asset to any company.
Vijay impressed me with his positive attitude and strong work ethic. Always eager to learn, with great initiative on challenging projects. His technical skills are equally impressive — it was clear from the very beginning that he had a great deal of potential.
Vijay's technical knowledge, attention to detail, and problem-solving skills are unmatched. He was always a key contributor to our team's projects. I highly recommend Vijay for any opportunity he pursues.
I am actively interviewing for senior Backend, Fullstack, AI, SRE, and Platform roles — remote, relocation, or Bengaluru. If you have an interesting distributed-systems problem, a team that moves quickly without cutting corners on quality, or you would simply like to discuss Haskell or Redis over a call, any of the addresses below will reach me.
fin. — thanks for reading the whole notebook.