I architect and build scalable products from the ground up: ticketing infrastructure, government digital platforms, and the backend systems that hold them together under load.
I like the hard, unglamorous middle (the schemas, the queues, the auth, the concurrency), and I care about how the whole thing looks and feels when it ships. I've built production systems end to end: a ticketing platform that sold 40,000+ tickets for international football matches, a statewide digital-services platform for the Government of Meghalaya, and CMS tooling used by officials across the state.
Outside of work, I keep a lab of experiments (implementing machine learning from scratch, biomimetic robotics, and local-first architectures), and I write them up with the graphs, the failures, and the results.
Full-stack ticketing platform for two international football matches: microservices for booking, payments, QR generation and admin, built to hold concurrency on sale day and feed a real-time stadium attendance dashboard.
Architected the backend and internal communication layer for a state service marketplace: robust REST APIs for interoperability, scalable PostgreSQL schemas for high-volume provider data, and DPDP-grade auth with encryption & hashing.
Independently designed, built and deployed a full-scale event ticketing & data-analysis system with admin and organizer modules: real-time booking and QR validation for everything from small meetups to stadium-scale operations. Later incubated by IIM Calcutta Innovation Park.
An interactive telemetry console showing how an in-memory cache shields a database through a 19,000-ticket on-sale. Swap write policies live, fire traffic spikes, and watch the system hold its line or melt down. Every metric is derived from an M/M/1 queue, not faked.
A load-test rig that fires thousands of simultaneous bookings at a mock seat-inventory service to watch row locks, queues and Redis hold-times behave under pressure.
An interactive simulator for tuning a line-following robot's controller: drag the gains, watch the error settle. Written up in detail on the blog.
Rebuilding the Onlybees analytics engine from a slow, racy sales ledger into a buyer-intelligence system: atomic writes, materialized rollups, cohort and RFM depth, and a recommend-only demand forecast.
A load-test sandbox fires a thundering herd at four seat-claim strategies, measures the oversells centrally, and replays the race seat by seat. Naive double-books, but one atomic Lua script holds the line.
A bio-inspired underwater robot, from SolidWorks sketch to 80-metre depth control: thunniform CAD, ANSYS CFD with a stall threshold, three control laws head to head, and where machine learning takes it next.