2026-10-01
Rate-limiting a shared quota from a stateless Cloudflare Worker, with no queue service, no lock, and no in-memory bucket. Keep one cursor for the next allowed time, turn waiting into a client-side reservation, cap under the real limit, and keep the decision in a pure function.
2026-09-28
The bottleneck is never the thing on the pricing page
I budgeted tokens for the AI feature and nearly blew the free tier on database writes. Scale every quota against the same load, then batch before you reach for a bigger plan.
2026-09-27
Let the model choose from a menu, never write the menu
The single constraint that turned a scary LLM feature calm — let it select from a curated set instead of generating freely — plus the validator and deterministic fallback behind it.
2026-09-26
I built an AI coach for zero rupees a month, and the LLM was the least of my problems
A real AI feature on free tiers alone. The model was the easy part; the thing that nearly broke the budget was not tokens at all.