Notes from building LLM systems for the enterprise: agents, retrieval, evaluation and LLMOps.https://matteocalzametre.com/blogenhttps://matteocalzametre.com/blog/ai-agent-observability-langfusehttps://matteocalzametre.com/blog/ai-agent-observability-langfuseWhat AI agent observability is, why agentic components need it, what tracing a modern agent actually means, and how I traced an on-premise enterprise agent with self-hosted Langfuse and LangGraph.Thu, 22 Oct 2026 00:00:00 GMThttps://matteocalzametre.com/blog/evaluating-agentic-text-to-sql-systemshttps://matteocalzametre.com/blog/evaluating-agentic-text-to-sql-systemsExecution accuracy gives one number to very different failures. How I evaluated an agentic text-to-SQL system in production with independent metric families: retrieval, output and SQL components, never folded into a single score.Mon, 19 Oct 2026 00:00:00 GMThttps://matteocalzametre.com/blog/validating-llm-generated-sqlhttps://matteocalzametre.com/blog/validating-llm-generated-sqlA system prompt is a request, not a contract. Here's the deterministic guard that parses every query an agent writes and refuses anything that isn't a safe read, before it reaches a production ERP.Thu, 08 Oct 2026 00:00:00 GMT