Notes on what I build, and what breaks along the way.

We moved a 1.5B-document OpenSearch collection off Serverless onto a single self-managed node, and cut the monthly bill from $1,800 to around $560. Here's the shard math that made that safe.

The biggest outages of 2025 and 2026 weren't attacks. They were config changes, cascades, and now AI agents: the same self-inflicted failures on repeat. Here's the pattern, and how to shrink your blast radius.

In 2017 a team made two different PDFs with the same SHA-1 hash, which is supposed to be impossible. But the headline gets the story backwards: that wasn't the death of SHA-1. It was the funeral, and the body still isn't buried.

Everyone says they're building AI agents. Few have shipped one. After building agentic features into a product of my own, here's why I'm convinced this is the direction, and why the hard part isn't the model.

I was pulling raw rows out of OpenSearch and computing facet counts in Node. Moving that work into the engine with terms aggregations cut average query time by 72%.