Writing
Notes on full-stack engineering, system design, and lessons from shipping production software.
A code review that only says "looks good" or lists nitpicks teaches nothing. After mentoring several junior engineers, here's what actually moves the needle.
Three payment gateways, three different reliability profiles, one idempotency pattern that made all of them safe to retry.
Most Node.js performance problems aren't about the language — they're about the query, the cache, and the thing blocking the event loop. A practical checklist.
AI-powered support isn't about replacing agents — it's about triaging correctly so agents spend their time on the 20% of tickets that actually need a human.
How to keep stock counts accurate to the unit across 10+ marketplaces selling the same SKU at the same time — without ever overselling.
Most systems I've built use both, side by side, on purpose. The decision isn't "which is better" — it's which data actually needs relational integrity.
A message queue without a dead-letter strategy just fails silently instead of loudly. Here's the retry and DLQ setup that's kept burst order volumes from dropping messages.
Neither one is universally better. The real question is which failure mode you'd rather deal with — REST's over-fetching, or GraphQL's query complexity risk.
REST is great for request/response. It falls apart the moment one action needs to trigger five unrelated side effects — here's when and how to reach for events instead.