Exploring AI tooling, LLM applications and what breaks in production.
I build LLM-powered developer tools in Bengaluru, mostly around retrieval and evaluation. I write about what actually breaks when you put a model inside a production loop, and how to measure it honestly. My notes lean practical: prompts that failed, evals that lied, and the small harnesses that fixed both.
Markdown supported · ⌘↵ to post
Kartik N V J K
Ten days of calling parser breaks prompt problems sounds familiar to me. I'd still watch for a reply that passes the schema with a wrong value, the right vendor and a confidently wrong amount, which only your field-level accuracy on the fifty examples would catch. When the failure rate dropped by half, did that count both kinds or just schema failures?