4/ Set your own timeout. SDK defaults are tuned for someone else’s worst case — we were sitting on a 10-minute default. Dropped to 20s with 1 retry. Fail fast, retry once, move on.
1/ Made our receipt-scanning API 61% faster this week. No bigger model, no new servers. Just fixed things that were quietly wasting time. What actually moved the needle 🧵
3/ Trim your LLM prompt — but only the real duplication. We were describing the same rules twice: once in free text, once in the JSON schema’s field descriptions. Cutting the redundant copy shaved 28% off the prompt with zero accuracy tradeoff — the schema still enforces it.
wanted to do cool things with AI
have instead spent the morning updating environment variables and pasting GitHub secrets
the future is here and it needs me to confirm the endpoint URL starts with https://