Writing

Field notes

What breaks when an LLM system meets production, and the order to fix it in.Subscribe by RSS.

  • · LLMOps · Data Engineering · RAG · 8 min

    The half of your GenAI system nobody put on call

    Nobody's RAG system breaks — it quietly stops being right. The jobs nobody wrote down, three mechanisms that make them safe, and what to put on a dashboard.

  • · LLM Serving · Evaluation & Monitoring · Regulation & Governance · 11 min

    Temperature zero is not deterministic

    Same prompt, same model, different answer. The mechanism behind non-determinism, the arithmetic that kills agent pilots, and what to do about it.

  • · RAG · Search & Retrieval · Evaluation & Monitoring · 9 min

    Your RAG doesn't have a retrieval problem. It has a ranking problem.

    The document was in the index and the answer was still wrong. How to tell a retrieval failure from a ranking failure, what to fix first, and what each costs.

  • · LLMOps · Evaluation & Monitoring · 9 min

    The gates are the product, not the orchestration

    The orchestration in an AI content pipeline is the boring part. What decides whether text ships is a set of model-free functions with an exit code.