Utsav Rai
Backend engineer, 22, from Bihar. I build production agentic AI systems in Node and TypeScript — and I write down what breaks.
I'm a backend engineer. I write systems in Node and TypeScript, mostly around production AI — model routing, fallback chains, evaluation harnesses, and the unglamorous plumbing that keeps all of it from falling over at 2 AM.
22, from Bihar, India.
What I actually work on
Day to day that means things like moving inference off a hosted API and onto self-hosted open-weight models behind a gateway, then building the A/B machinery to prove the swap didn't quietly degrade quality — and the eval harnesses that catch it when it does.
A lot of it is latency work: streaming versus non-streaming, where the tail actually sits, which timeout is lying to you. It turns out the non-streaming path is usually the one that times out in production, which is not what you'd guess from a benchmark.
The rest is ordinary backend craft at a size where ordinary stops being ordinary — schema changes on tables with nine figures of rows, queues, caches, migrations, and the interesting ways each of them fails once the row count gets serious.
Why this site exists
Most engineering writing shows you the finished thing. It skips the part where someone believed something false for six hours.
That gap is the useful part. A tutorial tells you what to type. A mistake tells you which of your assumptions is load-bearing and wrong — and that's the thing that actually transfers, because the specific bug won't repeat but the bad instinct behind it will.
So: Learn From Mistakes is one post per real thing I got wrong. The wrong belief stated plainly, the mechanism explained properly rather than hand-waved, and the rule I follow now. One a week.
I'd rather be the person who publishes the outage than the person with a clean blog and a loud incident channel.
Side projects
I build a lot of small tools, usually to scratch one specific itch. A running theme: most of them run with no API key — they borrow my Claude Code subscription by shelling out to the CLI, so a personal agent costs nothing per run. Browser agents, growth automation, an Android test platform, a LeetCode tutor that refuses to give answers.
See projects.
Elsewhere
- X / Twitter — where most of this gets discussed
- GitHub — the code
If you want to talk about backend, inference infrastructure, or something you broke recently, X DMs are open.