I'm Kartik.

I spend most of my time inside Apache Pinot working on upserts and realtime ingestion across the full spectrum of data sources we support. It's the kind of codebase where a single atomicity regression surfaces as a P0 at 3 AM and suddenly you're locked in whether you like it or not. I write Java for a living and I genuinely enjoy it. Not in the Stockholm syndrome way people assume. I just think the JVM is one of the most underappreciated pieces of infrastructure ever built and distributed systems work on the JVM still hits different.

Before Pinot consumed most of my attention I was already drawn to the problems that live below the service layer. Latency budgets that refuse to add up. Replicas that are consistent but wrong. Indices that pass every test suite you throw at them and then go OOM the moment production traffic arrives. That's the work I find interesting. The layer where clean abstractions stop and real tradeoffs begin.

More recently I've been building AI agents and custom harnesses around them. Not the wrap an API and call it autonomous kind. Actual multi step systems that plan and execute and evaluate their own output and course correct. Most of this work doesn't survive long enough to have a git history. If something makes it to a blog post I consider that a minor miracle. The rest got framemogged by my own negligence.

This site is where I write about the things I actually learned not the things I wish I'd known. Deep dives and postmortems and technical notes and the occasional rabbit hole that turned out to be worth the hours.

If something here is wrong I'd genuinely like to know.