Ruby application built by 'Pri' at Arkency over two-to-three weeks as an internal Wikipedia for the company, then open-sourced (finished the day before wroclove.rb 2026) after convincing Andrzej Krzywda. Ingests raw text (e.g. YouTube transcripts) and, given an ontology (node kinds + allowed relations for a domain), uses an LLM to extract nodes and edges into a knowledge graph. Each ingestion is a single transcript with multiple extraction attempts, tracked with stats (tokens used, time, LLM calls, round-trips, cost). Full audit log of extractions. UI browses nodes by kind with all their edges. Event-sourced — the entire graph can be rebuilt from the events. Ships with an MCP server so any LLM can query the graph conversationally. Entity resolution is enabled by exposing three Ruby-LLM-style tools (including list_nodes_by_kind / search_nodes / get_node_edges) to the LLM during extraction so it can reuse existing entities. Demoed at wroclove.rb 2026 with a graph built from the last six editions of wroclove.rb; asking the MCP server the most trending subject at the previous conference returned 'SQLite'. Built with Ruby LLM; used bleeding-edge Claude Opus 4.7 which wasn't fully priced in Ruby LLM at talk time. This very knowledge graph is the live output.