Notes on making company AI actually work
Arguments and arithmetic, mostly about the unglamorous half of the problem: what your systems know, what it costs to tell a model about it, and how to keep a record you can stand behind.
- 13 October 2026StrategyAdoption4 min read
Your AI pilot did not stall because of the model
Most internal AI projects die somewhere between the demo and the second week. The model is almost never the reason, and the thing that is has a name.
Read → - 10 October 2026CostOperations4 min read
What it costs to paste your company into every prompt
Background context is the largest line on most AI bills, it is re-sent on every turn, and almost nobody measures it. Here is the arithmetic.
Read → - 6 October 2026MemoryArchitecture3 min read
A knowledge graph is not a better vector search
RAG finds passages that look like your question. A graph knows two records are the same customer. The difference shows on exactly the questions that matter.
Read → - 1 October 2026GovernanceMemory3 min read
When your AI gets it wrong, can you prove what it knew?
Most AI memory overwrites. That is fine until somebody asks why a decision was made in July, and the only honest answer is that the record no longer exists.
Read →
New posts and every shipped change land in the feed. Or read the changelog, which is where the detail goes.