Nomad is designed for total data custody. The following specification outlines the reference hardware and service orchestration required for autonomous operation.
Model: DeepSeek-R1-70B
Handles high-level reasoning, task decomposition, and "Chain of Thought" planning.
VRAM REQUIREMENT:
96GB+ Pool (e.g. Dual 48GB)
Model: Qwen-2.5-32B
Handles syntax generation, tool execution (SSH/SQL), and writing code to RAM.
VRAM REQUIREMENT:
48GB+ Pool
Model: BGE-M3 (SOTA)
Converts documentation and source code into vectors for RAG retrieval.
VRAM REQUIREMENT:
12GB+
These services provide compute power and API interfaces. They remain passive until instructed by an agent.
Python scripts running in background loops. They actively monitor databases to self-trigger work.
The Doer. Monitors the job queue. When a task appears, it hydrates the RAM workspace, calls the Engineer to execute tools, and commits results.
The Safety Net. Scans active projects. If a project stalls (zero tasks), it generates a "Scoping Task" to keep momentum alive.
The Librarian. Reads raw execution logs, synthesizes patterns, and updates Long-Term Memory (Context Summary).
User sends intent to Nomad App (8000). The App retrieves context from Qdrant and forwards the package to The Architect (8002). The Architect generates a plan.
The Architect saves the plan as a Task in the Application DB and queues a job. The Executive Agent picks up the job and creates a temporary sandbox in the RAM Drive.
The Executive orders The Engineer (8003) to write code. Before execution, the request passes through the Action Approval Service (8004) to check for forbidden operations.
Code is tested and committed to disk. Later, the Consolidation Agent reviews the logs and updates The Memory (8005), ensuring the system learns from success.