Chunking is a retrieval decision, not a preprocessing detail
Where the boundary falls decides whether a rule and its exception arrive together. The pipeline that cannot delete a withdrawn document is the failure that survives handover.
5 min read
Ai engineering
Escrow does not survive model deprecation
Source escrow assumes the artifact is the source. Where behaviour depends on a hosted model the provider retires, the deposited package no longer reproduces the system.
5 min read
Guardrails are an architecture, not a filter
Controls belong at the boundary that can enforce them. Authorisation happens in the downstream system, not in the model, and autonomy level sets how much each layer must carry.
5 min read
LLM-as-judge, and what calibrating it against humans actually requires
A judge is a measurement component with order sensitivity and no established authority. Swap the presentation order, and a judge that picks the first answer both times has produced two conflicting verdicts.
4 min read
Prompt injection is a tool-permission problem
The injection is not the vulnerability; the tool the injected instruction can reach is. Retrieval and fine-tuning do not mitigate it, and prototypes grant permissions generously.
6 min read
Retrieval, fine-tuning, prompting: which problem each one solves
Three intervention points, not three exclusive strategies — the original RAG paper already combined two of them. Diagnose the failure before choosing the technique.
5 min read
Run cost is the number the prototype hides
Build cost is paid once; inference, retrieval, review and monitoring scale with use. A cheaper model bill can raise cost per accepted outcome, and the run-cost owner is absent from the approval.
4 min read