eStreetz
Notes from
the workbench.
What we learn building AI products for ourselves and our clients. Written by the people doing the work.
Claude, GPT, Gemini. What we actually reach for, and why.
A working team's notes on model choice in late 2026. Less leaderboard, more "which one do we open when a client calls on a Tuesday".
We cut our Claude bill roughly in half. Here's the boring list.
No clever tricks. Prompt caching, a shorter CLAUDE.md, cheaper models for the grunt work, and the batch API for anything nobody is waiting on.
Your agent doesn't need 40 tools
We gave an agent every API endpoint we had. It got slower, more expensive and more confused. What we do now instead.
RAG is mostly a search problem
When a "chat with your documents" project gives bad answers, the model is usually fine. The search in front of it isn't.
The checklist we run before an AI feature ships
It fits on one screen. It has saved us from at least three embarrassing launches.
Cheap models at the edge. When Workers AI is enough.
Not every AI feature needs a frontier model. A few jobs we moved to open models on Cloudflare, and a few we moved back.