Insights
We publish two to four pieces a month. Every piece is written as question-form subheadings followed by a direct answer, so a person and an answer engine read it in the same units.
Target of 2-4 pieces a month, tracked as a fixed KPI
Six mechanisms that stop a RAG assistant inventing answers
Suppressing wrong answers comes before raising the answer rate. Here are the retrieval, threshold, refusal and citation settings we actually run.
Seven things to check before replacing a manufacturing ERP
ERP projects fail not because features are missing, but because the data cleansing scope and the approval lines were never settled before work began.
Company sites in the age of AI search - llms.txt and question-form H2
The first reader of your site is no longer a person. Here is how to structure it so an AI answer cites you.
Export document automation - how far HS classification can be left to a system
Where HS classification can be automated, and where a person has to carry the responsibility.
SpaceSJ RAG engine v2 - clause-level version awareness
We solved answers grounded in withdrawn clauses of revised documents by filtering validity at clause level.
Where people should stay in an automated process
Why attempts to automate everything fail, and how to decide where the approval gate belongs.
Insights are written in English and Korean. Where a piece cites Korean regulation, the Korean original is the governing text.