Projects

Distributed marketplace where buyers and sellers trade through replicated trader services backed by a shared warehouse. Components run as separate processes over TCP so failures and concurrency show up the way they do in real multi-service systems. Three experiments are conducted —

Caching: An eventually consistent ledger cache at the traders reduced warehouse RPC traffic 5.8× versus reading through on every buy — trading stale reads for load, with oversell rising when invalidations lagged demand.

Fault tolerance: Killing a trader mid-run left throughput flat at 40 goods/s. Heartbeat detection and idempotent retries moved in-flight buys to a live trader fast enough to hold the line, though supply was seller-capped in that configuration.

Ordering and recovery: An earlier single-coordinator phase ordered concurrent buys with leader election, vector clocks, and deterministic tie-breaking, snapshotting state so a new coordinator could resume mid-queue. Re-election cost 10.6ms average client latency against 0.9ms steady state.

Roommate matching and housing placement for students — students set lifestyle habits and roommate preferences, get ranked recommendations, form groups, and receive a dorm assignment.

Compatibility scores are computed ahead of time rather than at request time: profile updates trigger background jobs that rebuild cached per-user rankings in Redis, so match results are served from a precomputed set instead of scanning every candidate on each request. Groups are scored the same way against their combined preferences, with housing assignment run in a transaction so two groups can't claim the same unit.

Physics-informed neural networks for modeling head-on vehicle crash dynamics — predicting stress, velocity, and displacement during impact without waiting on full finite-element runs for every design tweak.

Brought simulation turnaround from 25-hour FEM jobs to near-instant inference while staying within 18% of analytical and FEM baselines — accurate enough to screen design variants early and reserve full simulation for the promising ones. Transfer learning across structural variants reduced how often models needed full retraining. Completed as B.Tech thesis work at IIT Bhubaneswar.

End-to-end web platform for Wissenaire, IIT Bhubaneswar's annual technical festival — event registrations, file submissions, and payment checkout in one place for organizers and participants.

Processed 200+ registrations and ~$1.2K in transactions through Instamojo. A companion campus-ambassador app used the Facebook Graph API to track shares and award engagement rewards, helping drive reach across campus networks.