Writing

Notes on building LLM products: evals, extraction, and agents.

First post coming soon: what recovering 76% of failed receipt scans taught me about LLM fallbacks.