“We told the AI not to do that in the system prompt.”
Prompts are not permissions. Two recent incidents show why agent boundaries belong in a policy layer the model cannot rewrite.
Read →
← Back to site
Writing
Notes on platforms, data, and practical AI
The same things I write about on LinkedIn, kept here for anyone who prefers reading it in one place.
Every model call in this pipeline goes through one function.
One wrapper around LiteLLM, no provider SDKs, no structured output and no tool calls — and what actually turned out to be the safety net.
Read →
You will never “keep up with technology” again.
Full coverage stopped being a realistic goal. What the job runs on now is a framework solid enough to classify anything new, fast.
Read →
The checklist works for one agent. The platform is needed after the second.
Agent design checklists describe one agent, built by one team — they break down the moment a second team ships one.
Read →