Srikanth

A working issuer-processor — the system behind a card that decides in real time whether to approve a purchase, moves the money without ever charging twice, and keeps books that cannot go out of balance. Built the way Marqeta or Lithic would build it, at the scale one engineer can run and test end to end.

The same problem I solved at redBus across 7 gateways, rebuilt as a clean, tested system. Ingests daily settlement files from simulated gateways, matches them against a double-entry ledger, and classifies every mismatch — missing, duplicate, amount, fee, timing — with a tested rule per class. Designed so that events from different rails, with different schemas and latencies, resolve to one transaction rather than three rows someone stitches together by hand.

Every payment system needs a reliable stream of what changed. This pipeline syncs Postgres transaction changes to a downstream transaction store without polling the source — about 6 seconds from commit to queryable in Delta Lake. Debezium reads row-level changes from the write-ahead log, Kafka carries them, and Flink handles deduplication, late-arriving events, and schema evolution.

My B.Tech final-year team project at JNTU Hyderabad: RFID tags in student ID cards triggered an Arduino reader at campus gates, logging each exit event and raising alerts. I wrote the reader logic and event handling in C. My first system where every tap had to be recorded exactly once.