Writing
Notes from building two systems: an assistant that answers from the EU AI Act and cites every sentence, and an agent that writes to Asana and HubSpot only after a person approves. Each article says what I measured and what you can reuse without hiring anyone.
New articles also arrive by RSS.
Human approval for AI agents that write to real tools
How to let an AI agent change Asana and HubSpot safely: the model drafts, code does the dates, a person approves one exact version, each write happens once.
Case study: AI service delivery coordinator
How to stop a RAG assistant inventing citations
A RAG assistant that cites sources can still invent them. How I verify every citation and every sentence in code, refuse when a check fails, and log why.
Case study: EU AI Act compliance assistant
When hybrid search scores worse than vector search
Adding BM25 to vector search cut my RAG assistant's recall@5 from 0.86 to 0.58 on 36 labelled questions. What I tried, what I shipped, and what to test first.
Case study: EU AI Act compliance assistant
Tell me what is eating your team’s week.
Send a short brief: what the work is, who does it now, and what it costs you. I reply within two working days with a rough scope, a price range and whether AI is the right tool for it at all.
Send a briefOr write to keel.agents@gmail.com.