Notes, in public, on how these systems are built and measured, for anyone who wants to learn.
How Agav guesses each expense, asks on purpose, and trains a calibrated model on the answers
How well an AI model answers questions through Agav's MCP connector, measured against a household with known answers