PG 01  Architecture Portfolio · 2025–2026

The supply chain that decides
before the crisis does.

An explainability-first, EU AI Act-compliant enterprise AI architecture for regulated manufacturers who can no longer afford the intelligence gap between their data and their decisions.

8
Intelligent
Modules
5
High-Risk AI Systems
EU AI Act Annex III
3
Intelligence Layers
Sense · Decide · Comply
0
Black-Box
Decisions
EU AI Act Compliant TOGAF 10 GDPR · MDR 2017/745 Google Cloud · Vertex AI ADK Multi-Agent · A2A XAI / SHAP ISO 13485 · ISO 14001 CSRD Scope 3 SAFe · HITL Gemini 2.0 Flash · 2.5 Pro BigQuery · Pub/Sub FDA 21 CFR Part 820

01Why Now

The architectural conditions for autonomous supply chain AI
are now in place.

Four things had to be true simultaneously for an autonomous supply chain to be architecturally viable: the regulatory framework had to be clear enough to design against, the tooling mature enough to deploy, and the data infrastructure fast enough to act on in real time. In 2024–25, all four converged.

FACTOR — 01
Regulatory Convergence

EU AI Act, MDR 2017/745, and CSRD all entered enforcement simultaneously in 2024–25. The compliance surface for a regulated manufacturer tripled in 18 months. Architecture must satisfy all three simultaneously — or it satisfies none. The forcing function that turns explainability from a nice-to-have into a contractual necessity has arrived.

EU AI Act · MDR · CSRD · 2024–25
FACTOR — 02
LLM Contract Intelligence

Gemini 2.5 Pro's 1M+ token context window makes full-corpus SAP Ariba contract analysis possible for the first time — 480 suppliers, years of contracts, a single inference pass. No chunking. No information loss. The boundary between structured enterprise data and unstructured contract documents, which blocked enterprise AI adoption for fifteen years, has collapsed.

Gemini 2.5 Pro · 1M+ Context · 2025
FACTOR — 03
A2A Commerce Protocol

Google ADK and the Agent-to-Agent (A2A) protocol enable procurement agents to transact directly with supplier agents — issuing verified purchase orders, receiving structured bids, and writing every step to an immutable audit log. This is the first architecturally sound, auditable, agent-native commerce layer for enterprise supply chains.

Google ADK · A2A Protocol · 2025
FACTOR — 04
Real-Time Supply Chain Data Fabric

BigQuery, Pub/Sub, and Vertex AI Feature Store compose into a real-time data fabric that enables sub-second supplier risk scoring from live financial filings, news streams, and ESG data. The gap between a supply chain event and an AI-driven response has collapsed from days to seconds. The quarterly spreadsheet is architecturally obsolete.

BigQuery · Feature Store · Pub/Sub

02The Anchor Problem

Anchor Client · MedDevice Industries GmbH

MedDevice Industries GmbH
HQ: Düsseldorf, Germany  ·  Revenue: €2.4B  ·  Employees: 6,800 across 22 countries
Products: Surgical robotics components · Implantable device sub-assemblies · Sterile disposables
Supply chain: 480 active suppliers · 34 countries · 14 manufacturing sites
Regulatory surface: EU AI Act · GDPR · MDR 2017/745 · ISO 13485 · FDA 21 CFR Part 820 · ISO 14001 · EU CSRD

* Figures are illustrative, derived from published industry benchmarks for mid-market Class II/III medical device manufacturers. All scenario data is hypothetical.

34%
Forecast Error Rate *
€180M
Annual Inventory Cost *
€60M
2024 Stockout Cost *
23d
Avg. NCR Resolution *
PAIN 01
Demand Signal Blindness

12-week Excel forecasts updated monthly. No multi-signal intelligence — no POS data integration, no clinical trial pipeline signals, no economic leading indicators. The supply chain responds to history, not the future. When a tier-2 Malaysian supplier went dark with 11 days notice in 2024, there was no early warning system to absorb the shock.

34% forecast error · €180M inventory cost · €60M stockout in 2024 *
PAIN 02
Supplier Risk is a Spreadsheet

480 suppliers. Annual questionnaire-based risk assessment. No real-time monitoring of financial health, geopolitical exposure, ESG compliance, or sub-tier concentration risk. The supply chain has no awareness of the world its suppliers operate in until a crisis has already materialised. Risk arrives as a crisis — not as a signal with time to act.

Annual questionnaire cadence · Zero real-time financial or ESG monitoring
PAIN 03
Procurement is Manual and Fragmented

SAP Ariba deployed across four regional instances with different approval workflows, different contract terms, and different supplier master data. Strategic sourcing conducted without ML-assisted price benchmarking, risk-adjusted scoring, or contract clause analysis. Contract intelligence exists only inside individual procurement professionals — and leaves when they do.

4 Ariba instances · No ML sourcing · No contract intelligence layer
PAIN 04
Quality and Compliance Traceability is Reactive

NCR resolution requires manual trace across supplier certificates, inspection records, and batch records in three separate systems. Average NCR resolution time: 23 days. MDR Article 87 vigilance reporting SLA: 72 hours. The gap between what the architecture delivers and what the regulator requires is not a process problem — it is a structural liability.

23-day NCR resolution vs. 72-hour MDR Article 87 vigilance SLA
PAIN 05
Sustainability Reporting is a Manual Exercise

CSRD Scope 3 reporting required from FY2025. Emissions data lives in supplier questionnaires, logistics invoices, and freight export reports across 34 countries. Manual quarterly consolidation produces a CSRD report that is 90 days stale at the time of publication. The report describes where the supply chain was — not where it is.

CSRD Scope 3 mandatory FY2025 · Reports published 90 days out of date

03Design Philosophy

Four principles. Non-negotiable
in a regulated context.

These are not best-practice guidelines. In the EU AI Act, MDR 2017/745, and ISO 13485 regulatory environment, they are architectural constraints. Every component of this system must satisfy all four simultaneously.

PRINCIPLE — 01
Explainability engineered in — from model design to audit trail

XAI is not a dashboard added after the model ships. Every ML model in this system is designed with its explanation contract upfront — before a single line of training code is written. SHAP values are generated at inference time. Model Cards are versioned alongside models in the registry. Every ML decision produces a human-readable explanation before any write operation commits. When DemandIQ adjusts a replenishment order for surgical robotics components, the supply chain planner sees the top five demand signals that drove the adjustment, the confidence score, and a one-click override — before the SAP write occurs.

PRINCIPLE — 02
Human oversight is a first-class state machine node

EU AI Act Article 14 defines meaningful human oversight as a designed mechanism — a specific point in the decision flow where a named human reviews the agent's reasoning, the SHAP explanation, and the confidence score. In this architecture, every HITL checkpoint is a formal state in the agent state machine, with a defined entry condition, a presentation contract, a decision interface, a timeout behaviour, and an immutable audit record written before the agent proceeds. Article 14 is satisfied structurally, not by policy document.

PRINCIPLE — 03
Compliance obligations encoded as write-path constraints

EU AI Act, MDR Article 87, ISO 13485, and CSRD obligations are encoded as immutable constraints in the data model and enforced at write time. The compliance audit trail is a by-product of normal operations — not a separate process, not a monthly reconstruction exercise. Every event is tagged with the regulatory obligation it satisfies at the time of writing. A compliance audit reads the operational log. There is nothing to reconstruct after the fact.

PRINCIPLE — 04
Augment the enterprise — never replace its judgment

Every module has a defined autonomy boundary. Below the threshold: autonomous execution. Above the threshold: the agent prepares the best possible brief — supplier risk summary, SHAP explanation, confidence score, recommended action, and comparable precedents — and waits for a named human approver. Replacing human judgment in a medical device manufacturer is not the goal. Making it faster, better-informed, fully documented, and structurally compliant is.

04The Architecture

Four technical layers. Three intelligence layers.

The architecture has two complementary views. The technical stack comprises four layers (Experience, Agent Orchestration, MLOps, Infrastructure) — each with a single responsibility and clean interfaces. The intelligence model organises the eight capability modules into three layers: Sense, Decide, and Comply. Governance and compliance constraints flow downward through the technical stack; intelligence signals flow upward from MLOps to the Experience layer.

This is a concept overview. The full technical design — GCP reference architecture, Terraform IaC, ADK agent topology, and Vertex AI pipeline specifications — is developed in TOGAF ADM Phases A–F (PG 04) and the Infrastructure page (PG 08).

LAYER 01
Experience &
Presentation
8 Module Dashboards HITL Approval Surfaces XAI Explanation Viewer Supply Chain Command Dashboard Audit Trail Viewer
React · TypeScript
LAYER 02
Agent
Orchestration
ADK Multi-Agent Swarm A2A Commerce Protocol MCP Tool Manifest Pub/Sub Orchestration Agent State Machines Autonomy Boundaries Intra-Enterprise A2A Inter-Enterprise A2A
Google ADK · A2A Protocol · Cloud Pub/Sub
LAYER 03
MLOps &
Intelligence
Vertex AI Pipelines Feature Store SHAP · XAI Layer Model Registry Drift Detection Model Cards Confidence Scoring
Vertex AI · BigQuery · Gemini 2.0 Flash · Gemini 2.5 Pro
5 EU AI Act High-Risk Models (Annex III): DemandIQ · SupplierSentinel · QualityTrace · InventoryOrchestrator · ContractIntelligence
EU AI Act Annex III Classification Rationale
DemandIQAnnex III §8 — AI in supply chains for critical infrastructure components (Class II/III medical devices)
SupplierSentinelAnnex III §8 — Autonomous risk assessment affecting continuity of critical medical device supply
QualityTraceAnnex III §5(b) — AI used in safety components of medical devices subject to MDR 2017/745
InventoryOrchestratorAnnex III §8 — Autonomous allocation decisions affecting availability of life-critical device components
ContractIntelligenceAnnex III §8 — AI system materially influencing procurement decisions in critical supply chains
LAYER 04
Infrastructure &
Governance
Terraform IaC GKE · Cloud Run VPC-SC Zero-Trust CMEK Encryption BeyondCorp IAM · Workload Identity Immutable Audit Log Cloud Build CI/CD GreenOps · Carbon Footprint API FinOps Cost Tags
GCP europe-west3 · Terraform · Security Command Center · Cloud Carbon Footprint

05Capability Layers & Modules

Eight modules. Three intelligence layers.

The three intelligence layers — Sense, Decide, Comply — organise the eight modules by architectural role, independent of the four-layer technical stack. Intelligence flows upward: Sense modules feed signals to Decide modules; Decide outcomes feed Comply modules with the documented audit record.

Sense — Intelligence Inputs
DemandIQ

Multi-signal ML demand forecasting. Replaces 12-week Excel with real-time POS, clinical pipeline, economic indicator, and historical demand fusion. SHAP-attributed. Confidence-scored. EU AI Act Annex III §8 high-risk.

SupplierSentinel

Real-time supplier risk monitoring across 480 suppliers. Financial distress signals, geopolitical exposure scoring, ESG compliance monitoring, sub-tier concentration mapping. 30-day advance warning — not 11-day crisis response. EU AI Act Annex III §8 high-risk.

Decide — Decision Layer
ProcureGuard

ML-assisted strategic sourcing. Risk-adjusted supplier scoring, price benchmarking, multi-criteria evaluation. A2A commerce integration for autonomous purchase order issuance within defined autonomy boundaries.

ContractIntelligence

Full-corpus SAP Ariba contract analysis using Gemini 2.5 Pro 1M+ context window. 480 suppliers, full contract portfolio, single inference pass. Clause extraction, risk scoring, non-standard term flagging, HITL routing. EU AI Act Annex III §8 high-risk.

InventoryOrchestrator

Real-time inventory positioning across 14 manufacturing sites. Multi-echelon optimisation driven by DemandIQ signals and SupplierSentinel risk scores. Autonomous rebalancing within threshold; HITL for above-threshold moves. EU AI Act Annex III §8 high-risk.

Comply — Compliance & Reporting
QualityTrace

Device lineage tracing from supplier certificate through production batch to implanted device. Atomic write to ISO 13485 device history record. MDR Article 87 vigilance reporting at 72-hour SLA. NCR resolution from 23 days to same-day. EU AI Act Annex III §5(b) high-risk.

ScopeTracer

CSRD Scope 3 data fabric. Supplier emissions data collection, logistics invoice parsing, freight export integration. Real-time consolidation replaces quarterly manual process. ISO 14001 environmental data requirements satisfied via Cloud Carbon Footprint API integration.

Supply Chain Command

Unified executive visibility layer. Real-time supply chain health dashboard. Cross-module KPI aggregation, risk surface map, regulatory obligation status, pending HITL queue, immutable audit log viewer. The operating cockpit.

06A2A Commerce Protocol

From risk signal to executed
purchase order — without a form.

The A2A Commerce Protocol enables MedDevice procurement agents to transact directly with supplier agents — issuing verified purchase orders, receiving structured bids, and writing every step to an immutable audit log. This is the complete inter-enterprise agentic commerce loop.

▲  Trigger Event

SupplierSentinel detects a risk spike on the tier-2 Malaysian PCB supplier — financial distress signal combined with a regional geopolitical event. Risk score crosses the autonomy threshold. HITL fires. Procurement team reviews the SHAP-attributed risk brief and approves emergency re-sourcing. The A2A Commerce loop initiates.

MedDevice Procurement Agent
Broadcast sourcing request over A2A

The MedDevice Procurement Agent generates a qualified sourcing request — schema-validated, SHAP-attributed risk justification, HITL-authorised reference included. The request is broadcast to three pre-qualified supplier agents over the A2A protocol. No form. No email. No manual RFQ process. The message is a structured, signed, machine-readable commercial document.

Supplier Agents A, B, C
Evaluate capacity and return structured bids

Three pre-qualified supplier agents evaluate the sourcing request against current production capacity, lead time availability, and pricing constraints. Each returns a structured bid over A2A — schema-validated, machine-readable, with full provenance. The bid format is defined by the A2A commerce schema; no supplier-specific interpretation required.

ContractIntelligence Agent
Evaluate bids, surface recommendation with SHAP explanation

ContractIntelligence evaluates the three bids using risk-adjusted total cost of ownership scoring — incorporating SupplierSentinel's live risk scores, Gemini 2.5 Pro contract analysis of existing supplier terms, and delivery reliability history. The recommendation is surfaced to the CPO with a full SHAP explanation identifying the five factors that differentiated the winning bid.

CPO → HITL → A2A Execution
Approve, execute, audit

The CPO reviews the recommendation, the SHAP explanation, and approves via the HITL surface. The winning supplier agent receives a signed purchase order over A2A. The full transaction — from risk signal through bid evaluation to executed purchase order — is written to the immutable audit log. Every step has a timestamp, an approver identity, and a SHAP attribution.

This is the complete inter-enterprise A2A Commerce loop — from risk signal to executed purchase order — without a human touching a form, without an email chain, without a manual RFQ, and without a single unaudited decision. The audit trail writes itself. The compliance record is a by-product of the commerce layer working as designed.

07The Architecture Speaks to Every Seat at the Table

What this architecture delivers
— by leadership role.

VP Supply Chain / COO
Operational intelligence that anticipates — not reacts.

34% forecast error. €180M inventory cost. €60M stockout when a Malaysian supplier went dark with 11 days notice. DemandIQ replaces the Excel forecast with a multi-signal ML model that reads POS data, clinical pipeline signals, and economic indicators in real time. InventoryOrchestrator positions stock across 14 sites dynamically. SupplierSentinel gives you 30-day advance warning — not an 11-day crisis.

CCO / Chief Compliance Officer
Compliance as a structural property — not a reporting exercise.

Five high-risk AI systems under EU AI Act Annex III, each classified and justified by design. MDR Article 87 vigilance SLA of 72 hours versus a 23-day NCR resolution baseline. CSRD Scope 3 reporting required from FY2025. FDA 21 CFR Part 820 design control and DHR requirements satisfied via QualityTrace atomic writes. Every high-risk inference routes through a named approver with a SHAP explanation before any write operation commits. The audit trail writes itself during normal operations.

CPO / Chief Procurement Officer
Strategic sourcing with real intelligence — and a commerce layer that executes.

480 suppliers. Annual questionnaire risk. Four SAP Ariba instances. Strategic sourcing without ML-assisted benchmarking. ProcureGuard and ContractIntelligence give you real-time risk-adjusted supplier scoring, full-corpus contract analysis across the entire supplier base in a single inference pass, and — for the first time — a procurement agent that can issue a verified purchase order to a supplier agent over A2A, autonomously, with a full audit trail.

08Regulatory Grounding

Compliance is not a layer.
It is the foundation.

Each regulation imposes specific architectural constraints — not just documentation requirements. The design satisfies them structurally, at write time, as a by-product of normal operations.

Regulation Architectural Constraint Imposed Risk Level
EU AI Act — Annex IIIHigh-Risk AI Systems
Five high-risk AI systems classified per Annex III §5(b) and §8. Full SHAP explanation + named HITL approver + versioned Model Card + risk management documentation. Write-path enforcement. Every inference produces a human-readable explanation before any write operation commits.
High Risk
MDR 2017/745Art. 87–89 Vigilance
Device lineage tracing and vigilance reporting infrastructure. 72-hour Article 87 SLA satisfied architecturally via QualityTrace real-time NCR pipeline. Atomic write to ISO 13485 device history record at every supply chain event.
High Risk
FDA 21 CFR Part 820Quality System Regulation
Design control records (§820.30), Device History Records (§820.184), and CAPA documentation (§820.100) satisfied via QualityTrace atomic writes to the immutable data fabric. Supplier qualification records (§820.50) maintained in real-time by SupplierSentinel. Post-market surveillance data feeds ML retraining pipeline.
High Risk
GDPR / EU DPDPData Protection
Supplier PII confined within VPC-SC perimeter. CMEK encryption — MedDevice holds the keys. Data residency enforced: europe-west3. Right to erasure preserved via Firestore document-level deletion with audit record preservation.
Native to Infra
EU CSRDSustainability Reporting
ScopeTracer produces a real-time Scope 3 data fabric and CSRD-compliant reporting pipeline. Replaces manual quarterly consolidation. Report reflects current state — not 90-day-old data. Mandatory from FY2025.
Moderate
ISO 13485:2016Medical Devices QMS
Device batch records, supplier qualification records, and NCR documentation stored in immutable data fabric. Atomic writes to device history records at every relevant supply chain event. Post-market surveillance data feeds ML retraining pipeline.
Moderate
ISO 14001Environmental Management
ScopeTracer and GreenOps scheduling satisfy ISO 14001 environmental management data requirements structurally. GreenOps scheduling shifts non-time-critical compute to low-carbon grid windows via GCP Cloud Carbon Footprint API. Carbon attribution per inference run computed and logged.
Native to Infra

09The Design

The design starts here.

The following pages take every principle above and express it as concrete architecture artifacts — client requirements, TOGAF phases, workflow simulations, agent topology, ML pipeline design, and infrastructure code. Each page is independently readable. Together, they form a complete enterprise AI solutions architecture portfolio.