Insights & ideas.
Notes from building production machine learning — what works, what fails, and the decisions that actually matter. Written for engineers and the people who fund them.
RAG or fine-tuning? Start with neither
The question gets framed as a choice between two techniques. It is usually a sequencing question, and the correct first step is neither of them.
Read the postThe LLM bill nobody forecast: seven levers for cutting inference cost
Training ends. Inference bills on every request, for as long as the product exists. Here are the levers that actually reduce it, in the order we reach for them.
Read the postYour RAG system isn't bad at generating. It's bad at retrieving.
When a retrieval-augmented system gives wrong answers, the instinct is to blame the model or upgrade it. Usually the correct passage never reached the prompt at all.
Read the postWhy AI projects fail — and the four questions that predict it
Very few AI projects fail because the model was not accurate enough. They fail on problem framing or on deployment — and the modelling in between goes fine.
Read the postLet's build something
intelligent.
Tell us what you're trying to build — we'll tell you honestly whether AI is the right tool, and what it would take.
Start the conversationor write to us at [email protected]