Architecture
Layers, flows, and components of the AI Fabrick (Overseer, Router, Proctor, Workers, Auditor).
Architecture
Layers
- Overseer (Orchestrator)
- Plans execution, manages continuations, routes tasks.
- Router
- Decides whether to run
local(light) orheavy(RunPod/Docker).
- Decides whether to run
- Proctor
- QA checks: schema validation, safety rules, acceptance tests, thresholds.
- Workers
- Light = provider wrappers (OpenAI/Gemini/DeepSeek) in Next.js.
- Heavy = RunPod/Modal containers for GPU/long pipelines.
- 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 + resultsDecision 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