Selected writing
Practical notes on RAG evaluation, decision-support design, and the cost discipline that decides whether internal AI systems hold up in production.
2026-07-07
When I ran an enterprise AI procurement this June, the model shortlist took 5 minutes. The procurement questions took 5 hours. The useful lesson was not the model ranking; it was the procurement path around it.
Read the post →2026-06-24
Frontier pricing isn't your problem. Mismatched tasks are. Five cost-competitive models cover the 80% of internal AI tasks at 3x to 45x lower cost.
Read the post →2026-06-17
The model was rarely the problem in 4 RAG engagements. The 3 operational gaps — eval, retrieval, cost — decided which pilots reached production.
Read the post →