// PLATFORM PROTOCOL: BOOK-OUTLINE.MD AND AIPLUMBER SITE V1 BRIEF.MD COMPLIANT

Infrastructure
Before
Intelligence.

Stop deploying models into operational chaos. The model is a cheap commodity you rent; the plumbing is the asset you own. If your AI systems have no stop buttons or clear logs, they will leak money.

// PROOF, NOT THEORY // 4 REGULATED DEPLOYMENTS

Every rule, anchored to a real deployment

Most governance frameworks have zero real deployments behind them. Each safety rule here is proven in production, under a real regulator.

• Najm Insurance (SAMA) — vision AI, 40+ cities, 6,000+ daily cases

• Government of India / NMML — €10M+, 180+ endpoints, 99.9% SLA

• De Lijn (Belgium) — EU AI Act roadmap, board-approved

• US Restaurant Intelligence — 200 FTE → 3 agents, 90% cost cut

// ARCHITECTURAL OWNERSHIP

Who Controls the Plant?

Our infrastructure maps directly to defined enterprise leadership roles.

ROLE ALPHA

Plant Manager

Keeps the work flowing. Tracks prompt performance, model speed, and looks for bottlenecks in your daily workflows.

ROLE BETA

Chief Risk Officer

Prevents data leaks and sets safety limits. Uses automated stop buttons to freeze runaway AI loops before they cost too much money.

ROLE GAMMA

Enterprise Architect

Connects AI with your old databases. Uses simple, clean APIs instead of fragile, messy workarounds.

// THE COST OF UNGOVERNED AI

Every ungoverned loop burns tokens.

A single runaway agent re-querying itself can drain a month of API budget before anyone notices. No cache discipline, no kill threshold, no attribution — just tokens leaking into a bill. Governance is cost control: the same pipes that keep you compliant stop the meter from running away.

<30s trace any token spend to its cause (Pipe 2)

AUTO suspend on cost-threshold breach (Pipe 4)

CACHE repeat context served, not re-billed (Pipe 2)

ZERO silent budget surprises

THE AI EXHAUST PROBLEM

Model Collapse:
The 4-Step Check for AI Verification.

AI models are increasingly trained on AI-generated text. They get lazy, repeat themselves, and invent facts — the slow rot known as model collapse. Our 4-step check catches it before it reaches your clients.

// GROUND TRUTH AUDITS · NAJM INSURANCE

Measuring the Fracture Point

At Najm Insurance, we audited AI across 6,000 daily insurance cases in 40 cities. We ran identical prompts through different models and discovered that they disagreed on critical decisions.

DRIFT CROSS-ENGINE DIVERGENCE

Identical inputs can diverge across Claude, GPT, and Gemini. Without a filter, you ship unchecked mistakes straight into a regulated process.

The 4-Step Verification Plan

A simple verification plan to catch lazy or hallucinated AI outputs.

STAGE 01

Double-Check (Run on Two Models)

Flush your task through at least two different model brands. If they give the same answer, it is safe. If they disagree by more than 15%, stop the process and alert a human.

VERDICT: AUTO-STOP ON > 15% DIVERGENCE

STAGE 02

Stress-Test (Ask Hard Questions)

Stop models from giving generic, lazy answers. Force the AI to defend its logic and isolate its own assumptions instead of repeating safe generalities.

RUN ON: ALL REGULATED COMPLIANCE WRITES

STAGE 03

Anchor (Verify Source Material)

Never trust the model's memory. Force it to cite uploaded files. If the model relies on general training weights rather than your uploaded sources, flag it as an **[External Weight]**.

ESCALATE: ANY EXTERNAL WEIGHT IN TIER 3 SUBMISSIONS

STAGE 04

Human Check (Match Risk to Effort)

Match your human review to the danger of mistakes. Low-risk drafts get a quick read; high-risk financial submissions get line-by-line validation.

DISCIPLINE: PRE-DEFINE TIERING MATRIX

// CONTROL CONSOLE: RISK SIMULATION

Active Infrastructure Board

Test the physical safety valves. Simulate a standard runaway cost loop, or trigger a **Divergence Error** to see how the system halts before delivering bad outputs.

System Flow Map

// TELEMETRY INTERFACE V1.0

INTELLIGENCE IS RENTED · PIPING IS OWNED SYSTEM NORMAL // EGRESS OPEN

// MANUAL INTERCEPT VALVES

Isolation Control

FLOW RATIO: 1.0X

GATE VALVE: ARMED // ARD-3

Live Audit Logging

[00:01:02] Loading 6-Pipes topology...

[00:01:05] Checking SAMA & GDPR compliance parameters...

[00:01:08] System handshakes established on all 6 pipes.

[00:01:10] Egress open. Context pipelines operating normally.

// COMPLIANCE BY DESIGN SCHEMA

The 6 Pipes

We isolate model processes inside strict containment layers. This shields your core business databases from AI hallucination and runaway execution costs.

01 //

Scoped Accounts

No AI agent inherits broad admin privileges. Every system has strict database limits with zero default write permissions.

// COMPLIANCE LEVEL: SEC-1

02 //

Clear Logs

Every single decision, database query, and citation is logged on a read-only ledger. You can trace any AI action in under 30 seconds.

// COMPLIANCE LEVEL: AUD-1

03 //

Safety Stop Limits

Continuous monitoring evaluates speed and error volume. If the AI begins looping or spends too much, it suspends automatically.

// COMPLIANCE LEVEL: LMT-1

04 //

Human Sign-Off

High-stakes actions pause immediately and wait for manual approval. The AI cannot execute financial or legal tasks on its own.

// COMPLIANCE LEVEL: HITL-3

05 //

Safe Connections

Old core mainframes are protected behind strict gateway layers. The AI works with insulated data feeds, never direct core databases.

// COMPLIANCE LEVEL: ISO-2

06 //

Fixes & Records

Every single workflow error and human override is documented. This updates our system's core rules and builds an evolving defense moat.

// COMPLIANCE LEVEL: GRC-4

// DESIGN ARTIFACT: THE 6 PIPES

THE 6 PIPES — GOVERNANCE TOPOLOGY
├── NON-NEGOTIABLES (Pipes 1-4)
│   ├── 1. Constrained Identities    -> least-privilege service accounts
│   ├── 2. Attributable Actions      -> logged, reversible, traceable <30s
│   ├── 3. Human-in-the-Loop Gates   -> high-stakes actions pause for a human
│   └── 4. Kill Threshold Monitoring -> telemetry + auto-suspend on breach
└── ENTERPRISE-FIT (Pipes 5-6)
    ├── 5. Legacy Insulation         -> API gateway wraps legacy systems
    └── 6. Learning Loop             -> ADR-driven evolution, not patching

// THE IP, IN LONG FORM

The AI Plumber

A governance-first field manual for shipping agentic AI into regulated production. The 6 Pipes, the 5-phase lifecycle, and four real regulated deployments — one per non-negotiable. No magic, just the plumbing the model doesn't give you.

// WHAT'S INSIDE

Part I — The Problem. Why ~70% of enterprise AI stalls at the governance layer, not the model.
Part II — The Framework. The 6 Pipes; the 4 Non-Negotiables in depth, each proven by one real case — Najm (SAMA), Government of India / NMML, De Lijn, US Restaurant Intelligence.
Part III — Implementation. The engagement playbook, the governance checklist, the three tiers.

// AVAILABLE NOW

Start with the whitepaper

The book is in final consolidation. The whitepaper distils the thesis and the 6 Pipes today — read it now, get the book when it ships.

Download the Whitepaper (PDF) The Book Landing →

// RECOVERY LEDGER & COMPLIANCE CAPEX INDEX

Financial Leakage Calculator

Unchecked model calls, duplicate prompts, and bugs waste significant capital. Adjust the sliders below to see your estimated annual savings.

Operating Parameters

// RUNTIME METRICS

Active AI Workflows / Pipelines 15

Monthly API / LLM Spend ($) $12,000

Weekly Developer Hours Spent Fixing / Patching AI 25 hrs

ANNUAL COMPLIANCE SUMMARY

// Recovered Operating Capital

$0

Prompt Caching Savings (30%): $0

Reclaimed Developer Hours (65%): $0

Divergence Error Avoidance: $31,500

Estimated Payback: 1.5 Months

// PROFILE SUMMARY

Koen Van Lysebetten

Enterprise Conversational AI Architect & Fractional CAIO

EMAIL: koen@devgap.uk

LINE: +32 469 41 31 52

HAACHT, BELGIUM

I have over 20 years of experience building and deploying complex language systems under strict regulatory limits. I served as one of the very first engineers at AutoScout24, and completed the successful scale-up of Kapaza to its Schibsted acquisition.

I manage DevGap, directing cross-border engineering teams across Europe, the Middle East, and India. We construct the structural boundaries that protect enterprise balance sheets from model hallucination and runaway API loop costs.

// STRUCTURAL CAPABILITY MATRIX

AI & Agentic Systems

· LangGraph orchestrations
· Conditional fallbacks
· Retrieval evaluation
· Dynamic routing layers

Cloud & Infra

· Docker & Kubernetes
· Azure DevOps / Git pipelines
· IaC Terraform configs
· Model registry schemas

Security & Observability

· EU AI Act compliance
· GDPR Article 9 protections
· Zero Trust architectures
· Prometheus & Dynatrace monitoring

C-Suite Translating

· Executive roadmapping
· Financial leak calculations
· Multi-vendor evaluation
· System ADR logs

// VERIFY & ENGAGE

WHITEPAPER v3.0 The Why Behind the Tech Governance-first architecture · EU AI Act · SAMA · GDPR

[↓ PDF]

FULL CAREER HISTORY LinkedIn / Koen Van Lysebetten 20+ yrs · DevGap · Najm/SAMA · Gov. of India · SMALS

[↗]

15-MINUTE GOVERNANCE REVIEW Book a Review Walk one agentic workflow against the 6 Pipes

[↗]

Infrastructure Before Intelligence.