AI Governance Unit (AIGU) — Continuous AI Governance Orchestration¶

TL;DR: A 9-agent LangGraph system that replaces manual AI governance checklists with a continuous orchestration loop. Projects flow through a structured lifecycle — Intake → Risk → POC → Pilot → Librarian → Gatekeeper → Production → Handover — with specialized Amazon Nova agents handling each stage, automatic delta detection, and a universal support overlay at every step.

Stack: Python • LangGraph • Amazon Nova (Pro/Lite) • AWS Bedrock • AWS Lambda • DynamoDB • S3 • React Native • Expo • LangFuse

✨ Features¶

🧠 9-Agent Brain - Each lifecycle stage has a dedicated agent with focused responsibilities; no monolithic prompt
⚡ Dual Paths - Hero capabilities fast-track through an Accelerator path (skip POC); New capabilities follow the full Standard path with CAF approval
📊 15% Delta Threshold - Production changes are auto-analysed; updates exceeding 15% trigger full GIGC re-approval automatically
🔄 State-Loop Architecture - React Native frontend is a "dumb" terminal; all state lives in DynamoDB, managed by LangGraph
🛡️ Universal Support Overlay - A read-only Support & Insights agent is available at every stage for status queries and "why" explanations
📋 Artifact Management - Governance Librarian validates required artifacts by risk level and prevents duplication via automated audits
🔍 LangFuse Tracing - Full prompt management and observability across all agent invocations
🧪 TDD Lifecycle Simulations - Pytest-driven end-to-end lifecycle tests covering all paths and edge cases

🧠 Architecture¶

graph TB
    subgraph Client_Layer
        UI[React Native / A2UI\nDumb terminal]
    end

    subgraph Logic_Layer
        Lambda[AWS Lambda\nAPI Handler]
        Graph[LangGraph\nState Machine]
    end

    subgraph Agent_Brain
        Intake[1. Intake Orchestrator\nRoute: Accelerator / Standard / Stop]
        Risk[4. Risk & Triage\nSLA: 3 / 7 / 10 days]
        POC[2. POC Agent\nNew capabilities only]
        Pilot[3. Pilot Agent]
        Lib[5. Governance Librarian\nArtifact validation]
        Gate[6. Gatekeeper\nGIGC admin approvals]
        Prod[7. Production Agent\n15% delta check]
        Handover[8. Handover Agent\nIRIS / LCT / RTB tasks]
        Support[9. Support & Insights\nRead-only, always available]
    end

    subgraph AWS
        Bedrock[[Amazon Bedrock\nNova Pro / Lite]]
        Dynamo[(DynamoDB\nGlobal State + Checkpoints)]
        S3[(S3\nArtifacts)]
        LF[[LangFuse\nTracing]]
    end

    UI <-->|A2UI protocol| Lambda
    Lambda --> Graph
    Graph --> Intake --> Risk --> Lib --> Gate --> Prod --> Handover
    Intake -->|New capability| POC --> Pilot --> Risk
    Gate -->|Blocked| Support
    Prod -->|Delta >15%| Gate
    Graph <--> Bedrock & Dynamo & S3 & LF

🎯 What Makes This Special¶

Governance as a State Machine, Not a Spreadsheet¶

Traditional AI governance is checklists, email chains, and manual sign-offs. AIGU models the entire lifecycle as a LangGraph state machine. Every transition — approval, rejection, delta re-review — is a graph edge. Stage-specific agents run in sequence; the Support agent is always available as a side channel. The governance process becomes auditable, automatable, and queryable.

Delta-Driven Re-Approval¶

When a production AI system changes, most governance processes require a full manual re-review by default. AIGU's Production agent computes the delta between the current and prior submission. Changes under 15% proceed without re-approval; changes over 15% automatically re-route through the Gatekeeper. The threshold is configurable and the logic is transparent.

Hero vs. New Paths¶

Not all AI capabilities need the same scrutiny. "Hero" capabilities (proven patterns within existing guardrails) skip the POC stage entirely and fast-track to Pilot. "New" capabilities follow the full Standard path including CAF approval. The Intake agent classifies the path from the initial submission; humans only intervene at defined gates.

🚀 Technical Highlights¶

LangGraph Orchestration¶

State machine: aigu/graph.py — a compiled LangGraph graph with conditional edges for path routing
State persistence: DynamoDB-backed LangGraph checkpoints for resumable sessions
Agent dispatch: each agent in aigu/agents/ receives the full state, acts on its slice, and returns a patch

Amazon Nova via Bedrock¶

Nova Pro: used for high-stakes decisions (risk classification, gatekeeper review, production delta analysis)
Nova Lite: used for lower-stakes parsing (intake extraction, librarian checks, support queries)
LangFuse: all prompts versioned and served at runtime; traces every invocation for debugging

Frontend (React Native / A2UI)¶

Pure state-renderer: the UI has no business logic; it renders whichever A2UI component tree the Lambda returns
expo start --web for browser preview during development

Deployment¶

./aigu_manager.sh deploy — full stack via CloudFormation (Lambda + DynamoDB + S3 + IAM)
./aigu_manager.sh teardown — clean removal of all AWS resources
pytest tests/simulate_full_lifecycle.py — end-to-end lifecycle simulation (no AWS required for unit stages)

📊 Key Metrics¶

Agents: 9 specialised agents covering the full AI project lifecycle
Lifecycle stages: 8 stages (Intake → Handover), 2 paths (Accelerator / Standard)
Delta threshold: configurable, default 15% — triggers automatic GIGC re-approval
SLA options: 3 / 7 / 10 days, risk-based assignment by Risk & Triage agent

This project demonstrates multi-agent orchestration for a real enterprise problem domain: replacing fragmented governance checklists with a continuous, auditable, agent-driven workflow.