I build the parts of a product that have to stay up.
For the last four years that has meant Praja — a regional social platform — where I've worked across a Ruby on Rails monolith, a set of Go and NestJS services around it, and the Kubernetes fleet all of it runs on. Along the way I've owned migrations, designed the release path, and shipped the infrastructure a paid tier depends on.
I like the problems where the answer is a measurement rather than an opinion: which queue is actually the bottleneck, what the p90 really is, whether the new model is better or just different.
Systems I designed, owned or rebuilt — described at the architecture level.
185 models, 211 background workers and ~28 Sidekiq queues serving a regional social network. Four years of feature work, performance work, and the discipline of keeping a large monolith habitable.
Helm charts, per-environment values and automated image bumps describing every service in production — most changes deliberately small, so the tree stays reviewable.
NestJS and Socket.IO with Redis pub/sub for cross-instance fan-out and MongoDB for persistence. Three years of ownership — including adding a message broker and later removing it.
A live migration from AKS to EKS — Terraform for the new estate, one service at a time, DNS and CDN last, then a deliberate decommission. No maintenance window.
Compositing user photography onto templated poster and video designs at festival-day volumes. Moved from Lambda + FFmpeg to KEDA-scaled in-cluster workers driven by queue depth.
Background removal moved from per-call cloud API to an ONNX model on GPU nodes in-cluster, then a face-detector swap validated against 200 real production images before shipping.
Targeting users by district, state, party and circle, then dispatching push through FCM without touching the main API's request path.
CircleCI with timing-based test splitting, build capacity on our own cluster, plus the instrumentation and backup drills that make production legible.
An inverted index, BM25, Levenshtein automata, FSTs, BKD trees, a byte-level segment format and a breakable cluster — ~20,000 lines of TypeScript with a 43-chapter book.
A migration engine that parses Ruby with tree-sitter, lifts it into a typed IR, and scaffolds Goravel code — with static gates that refuse to publish a partial tree.
A Ruby gem that tracks in-flight Sidekiq jobs and re-enqueues whatever a killed worker was holding, because Kubernetes evicts pods and Sidekiq does not remember what it lost.
Apr 2025 – Present
Circleapp Online Services (Praja)
Platform and infrastructure ownership alongside product work: the Kubernetes fleet, the media generation pipeline, and the release path everything ships through.
May 2022 – Mar 2025
Circleapp Online Services (Praja)
Backend feature work on the Rails monolith and the services around it, growing into the multi-cloud migration and the platform work that came with it.
Dec 2021 – Apr 2022
Circleapp Online Services (Praja)
Started on the Flutter client, then moved toward the API that fed it.
2021
Continual Engine
Applied-ML internship on image understanding for accessibility tooling.
B.Tech, Computer Science
Rajiv Gandhi University of Knowledge Technologies · 2022 · GPA 9.3 / 10
Notes from the work, published on the blog.
Open to conversations about backend, platform and infrastructure work.