Hard-won lessons from building, evaluating and operating RAG systems, agents and generative pipelines. No hype — just what works when real users depend on it.
Full autonomy is rarely the goal. We share the decision framework we use to place humans exactly where they add the most safety and value — and nowhere else.
Read articleYou can optimise for any two, but the third always pushes back. Here is how we reason about the trade-offs and tune LLM apps that stay fast, cheap and good enough.
Read articlePrompts are code: they need version control, tests, reviews and CI. We explain how to treat them that way before they quietly become your biggest source of production risk.
Read articlePgvector, Qdrant, or a managed service? We cut through the benchmark noise with the questions that actually determine which vector store fits your workload.
Read articleWhen one agent isn't enough, coordination becomes the hard part. These are the orchestration patterns that survived contact with production for us.
Read articleA practical, layered approach to keeping generative systems on-brand, on-policy and out of trouble — without smothering the quality that made them worth building.
Read articleVibe checks don't scale. We break down the offline and online metrics that actually predict whether your LLM feature is helping or hurting real users.
Read articleGiving a model tools changes everything about your risk surface. Here is how we design tool-using agents that are capable, bounded and auditable.
Read articleChunking is the most under-rated lever in RAG. We compare the strategies we reach for and explain how to pick one from the structure of your own documents.
Read articleFine-tuning is seductive and usually the wrong first move. We lay out why retrieval wins for knowledge work, and the narrow cases where fine-tuning still earns its keep.
Read articleWe publish practical write-ups on production AI a few times a month. Want them first — or want to talk through a problem of your own?