Writing by Niraj Jha
Architecture decisions and the trade-offs behind them, retrieval that survives production, and the things I got wrong first. Written from systems that shipped, not from tutorials.
9 posts
I turned the network off, liked my own post, and watched the count go up anyway. The lie was deliberate and correct here. Where it stops being correct is the whole question.
A support bot answered ninety questions four thousand times and paid full price for every one. Exact-match caching hit two percent. Here is what actually works.
Better Auth hashes passwords with scrypt, which wants 32 MB of memory per login. On Cloudflare Workers that is a problem. Replacing it was the easy part.
The skeleton took an afternoon and nobody ever saw it. The file was in the right folder with the right name, and four separate things stopped it rendering.
The bug report was one line: it deletes twice sometimes. The cause was a button that looked dead for a moment, and the fix we wrote was React's own flag, badly.
A two line pull request nearly went through on my phone. It was the third delete button that month, written the third different way, and one of them was wide open.
A three line accessibility fix made navigation instant and the orders list slow. The obvious next move made it four times worse, and the reason is one default.
An interviewer asks how to add a column to a live 250 million row table. The obvious answer is correct, and it will still take your site down.
I build AI products end to end at Creatr: email pipelines on Gmail webhooks, M365 integrations, retrieval that survives production. What I have shipped, and four things I got wrong first.
Open source
I run an open-source project like a real engineering team: a board, cycles, proper code review and written feedback. You join as a team member, not as a student.
How it works