pentoai/ml-ralph

Automates the full ML experiment lifecycle—planning, execution, analysis, and structured learning accumulation—through a four-phase cognitive cycle (understand, strategize, execute, reflect) that prioritizes verification and hypothesis testing over raw execution. Integrates with Claude Code CLI and runs within tmux, providing a terminal UI that operates autonomously on ML projects once initialized with a product requirements document. Implements a "paranoid scientist" framework allocating 70% effort to data verification and assumption validation, enabling strategic retreats to understanding phases when experiments plateau.