I’m Krishan, an engineer, and this is where I write up my own work.

Almost everything I have learned at a keyboard, I learned from someone else’s write-up — a post by a developer who had already hit the bug I was staring at and wrote down what actually fixed it. This blog is my half of that exchange.

So it reads as a working log rather than a tutorial site. The earlier posts come from years of machine learning: BERT and transformers, PyTorch training and distributed runs, mixed precision, JAX, and the Python, Java and tooling notes that went alongside them. The recent ones are about building and operating systems of AI agents — handing work to an agent without blocking on it, telling a finished job from a merely claimed one, the gates a change has to pass before it counts as done, and keeping a codebase from growing just because generating code has become cheap.

What most of it has in common is that I got something wrong first, measured it, and then wrote down the version I wish I had read. Writing it down is also how I remember it.