Turojin/jarvis-mvp

J.A.R.V.I.S. — Alvin OS Personal AI Assistant (MVP)

A locally-run, multi-agent AI assistant built on the Alvin OS reasoning pipeline. J.A.R.V.I.S. reasons through a 12-step swarm-orchestrated pipeline, remembers every turn, and speaks with dry wit and absolute loyalty to Alvin.

Quick Start

# 1. Install dependencies
pip install -r requirements.txt

# 2. Ensure Ollama is running with a model (e.g., llama3.1)
ollama pull llama3.1

# 3. Run
python main.py

Architecture (12-Step Pipeline)

Step	Name	MVP Status
1	State Reader	✅ Implemented
2	Contract Classifier	✅ Implemented
3	Dial Setting	✅ Implemented
4	Archetype Blender	✅ Implemented
5	J.A.R.V.I.S. Persona Injection	✅ Implemented
6	Council of Three	✅ Implemented
7	Self-Scoring Agent	⏳ Stub (disabled in config)
8	Mirror Check (Drift Guard)	⏳ Stub (disabled in config)
9	[ANCHOR] Gate	⏳ Stub (disabled in config)
10	Memory Writer	✅ Implemented
11	Proactive Loop Detector	⏳ Stub (disabled in config)
12	Response Delivery	✅ Implemented

Tech Stack

• Orchestration: smolagents (Hugging Face)
• LLM Backend: Ollama (local) with optional OpenAI/Anthropic fallback
• Memory: ChromaDB (vector) + SQLite (structured)
• UI: Terminal (Rich), FastAPI+HTML later
• Language: Python 3.11+

Configuration

Edit config.yaml or set env vars:

llm:
  provider: ollama      # or openai / anthropic
  model: llama3.1
  base_url: http://localhost:11434/v1

Environment overrides:

• JARVIS_LLM_PROVIDER
• JARVIS_LLM_MODEL
• JARVIS_LLM_API_KEY
• JARVIS_LLM_BASE_URL

Project Structure

jarvis/
  __init__.py
  config.py       # YAML/JSON config dataclasses
  llm.py          # Pluggable LLM backends
  memory.py       # SQLite + ChromaDB unified memory
  pipeline.py     # 12-step Alvin OS pipeline
  ui.py           # Rich terminal interface
main.py           # Entry point
config.yaml       # Default configuration
requirements.txt  # Dependencies

Handoff Notes

• Single-user system: assumes the user is Alvin.
• Privacy-first: all data on-disk, no telemetry.
• Modular pipeline: each step is a function; easy to swap orchestration frameworks.
• Future sprints will add self-scoring, mirror check, proactive loops, and a web UI.