Guides
Engineering guides
Long-form notes on the parts of an AI system that decide whether it can be shipped: where a person approves, how retrieval is tuned, where memory lives, and what a document review has to produce. Every claim in these pages points at a file and a line in a repository.
Agents and approvals
- Human-in-the-loop approval for AI agents
Three places to put a person in an agent's loop, and the code that parks a run.
- Multi-agent orchestration is a routing problem
Why a second agent rarely helps, and how the routing layer is actually wired.
- Self-hosted AI agent: what you take on
What running your own agent means once the demo is over.
Retrieval and memory
- Hybrid retrieval, tuned
When BM25, when vectors, and when a reranker is a waste.
- GraphRAG in production: what the paper leaves out
Entity extraction, storage and the cost the paper does not mention.
- Agent memory: where it lives, when it expires, how it is deleted
Memory as files, vectors and expiry instead of a black box.
Documents and protocols
- AI document review that locates its findings
From a located clause to a clickable highlight, with the human gate.
- MCP server architecture
Tool ownership, credentials and the allowlist behind one process per session.