Turning Complex Infrastructure Into Elegant Automation

Self-hosted personal knowledge system (RAG). Ingests notes, code, documents, and infrastructure data from multiple sources, auto-categorizes with hybrid rule-based and LLM tagging, stores vector embeddings in pgvector, and makes everything queryable by any AI assistant via MCP or REST API.

Self-prompting agent system for Claude Code. Instead of a human re-typing prompts, the environment assembles the next instruction — the agentic loop itself is the harness. From a single prompt it grows a real, versioned project, verifies what is objectively checkable with an ungameable grader, honestly labels what is not, and surfaces to a human approval gate.

Pointer-first memory hub for a fleet of AI agents. Sessions recall cheap summaries with an addressable handle on every line, zoom a handle to the level of detail they actually need, and store typed memory objects — decisions, traps, open threads of work — that outlive the session which wrote them. Layered scopes resolve nearest-first, so a project memory shadows a global one without either being lost.