Guides for scoping and evaluating AI work
Written for the person who has to sign off on the engagement, and for the engineer who has to deliver it.
Writing an AI brief that gets serious proposals
The four things engineers look for first — the data, the constraints, the evaluation method and the deployment target — and how to state them without over-specifying the solution.
Success criteria that survive contact with production
Why 'improve accuracy' is unenforceable, and how to write criteria against a held-out set, a latency budget and a cost ceiling instead.
Structuring milestones for research-shaped work
Splitting exploratory phases from delivery phases so uncertainty is funded honestly and neither side is trapped by a fixed price on an unknown.
Reading proof of work without being an ML researcher
What a model card, a benchmark table and a repository's commit history actually tell you about whether someone has shipped.
Inference cost reduction: where the money usually is
Batching, quantisation, caching, routing and prompt shape — the sequence most teams should attempt before buying more GPUs.
A checklist before rebuilding your RAG pipeline
Chunking, hybrid retrieval, reranking and evaluation harnesses — how to tell which layer is actually failing.