Dojo - Small Language Model Playground

"Ninjas who can clone themselves"

A playground for composing small language models into powerful systems.

Philosophy

• Composition over Scale: Many small, sharp tools > one monolithic model
• Unix Philosophy for AI: Each ninja does ONE thing well
• Dialectic Thinking: Thesis → Antithesis → Synthesis
• Scrolls: Universal data primitive with lineage tracking

Structure

dojo/
├── scroll.py      # Universal Scroll primitive (any type, with lineage)
├── ninja.py       # Ninja agents (small model wrappers)
├── hokage.py      # Orchestrator that commands ninjas
└── README.md

Usage

# Run on Linux box with Ollama
python3 scroll.py      # Demo universal scrolls
python3 hokage.py      # Run the orchestrator

Concepts

Scroll

Inspired by micrograd's Value, but universal:

• Numbers add, multiply, backpropagate
• Strings concatenate
• Dicts merge
• Lists join
• Functions compose
• ALL track lineage

Ninja

A small model wrapper that:

• Specializes in one task
• Can clone itself for parallel work
• Composes with other ninjas

Hokage

The orchestrator that:

• Commands multiple ninjas
• Routes tasks to specialists
• Synthesizes results

Hardware

Designed to run on modest hardware:

• RTX 4060 8GB or similar
• 32GB RAM
• Ollama with qwen2.5:1.5b (~400ms inference)