Layers, flows, and components of the AI Fabrick (Overseer, Router, Proctor, Workers, Auditor).

Architecture

Layers

1. Overseer (Orchestrator)
   • Plans execution, manages continuations, routes tasks.
2. Router
   • Decides whether to run local (light) or heavy (RunPod/Docker).
3. Proctor
   • QA checks: schema validation, safety rules, acceptance tests, thresholds.
4. Workers
   • Light = provider wrappers (OpenAI/Gemini/DeepSeek) in Next.js.
   • Heavy = RunPod/Modal containers for GPU/long pipelines.
5. Auditor
   • Logs decisions/events to Redis + S3 for reproducibility.

Execution Flow

sequenceDiagram
    participant User
    participant Overseer
    participant Router
    participant Worker
    participant Proctor

    User->>Overseer: Submit Prompt
    Overseer->>Router: Decide tier
    Router-->>Overseer: local / heavy
    Overseer->>Worker: Run task (stream)
    Worker-->>Proctor: Output
    Proctor-->>Overseer: QA pass/fail
    Overseer-->>User: SSE events + results

Decision Policy

Signals scored by the Router:

• Duration forecast: > soft budget → heavy/continuation
• Capability needs: GPU, long-context, tools → heavy
• Safety risk: high → heavy proctor or human-in-loop
• Latency target: interactive → prefer local
• Cost ceiling: avoid overage
• Reliability: select environment with better quotas
• Data governance: PII/export constraints may force local