RAG7 min readPlaceholder
What actually breaks when RAG reaches production
Retrieval demos fail in ways that are invisible until real documents arrive. The three failures that account for most of it, and what to measure instead of vibes.
ReadWriting
Notes from the
part after the demo.
Most of what is hard about AI engineering happens after the thing works once. These are notes on that: retrieval that fails quietly, evaluation nobody budgeted for, and the architecture decisions that are cheap on paper and expensive in code.
Operations6 min readPlaceholder
The part after launch is the part you are paying for
Models drift, prompts rot, and upstream APIs change their output shape without telling you. What a maintenance loop for an AI system actually contains.
ReadPractice5 min readPlaceholder
Write the architecture down before you write the code
A page of prose costs an afternoon and settles the arguments that would otherwise be settled in a pull request three weeks later.
Read