writing
Engineering notes
Notes on building AI systems where retrieval, authorization, evaluation, and security matter as much as the model.
01/4 min read
Read →Why Grounded AI Should Be Allowed to Say “I Don’t Know”
Retrieval alone does not make generation trustworthy. Evidence sufficiency and abstention belong in the product contract.
RAG · Grounding · Evaluation
02/4 min read
Read →Tenant Isolation for AI Systems Belongs Below the Prompt Layer
Prompts are not a security boundary. Multi-tenant AI needs organisation and workspace isolation enforced in application code and the database.
Security · PostgreSQL · RLS
03/3 min read
Read →Evaluation as a Release Gate for AI Systems
“It seemed to work” is not a release decision. Repeatable evaluations turn AI behaviour into something you can regress-test.
Evaluation · Testing · AI Engineering
04/3 min read
Read →Giving AI Tools Without Giving It Uncontrolled Power
Tool-using AI should propose intent. Authorization, confirmation, and side effects belong to deterministic backend code.
AI Security · Authorization · Backend