Local Model Execution & Data Sovereignty

How AML Labs deploys agentic AI for compliance without transferring sensitive data to third-party ML providers, external APIs, or cross-border cloud regions — ensuring full regulatory control.

Core Principles

Why Local Execution Matters

Financial institutions operating under AML/KYC regulations face strict data residency and processing requirements. Transmitting customer PII, transaction records, or risk assessments to external ML inference APIs introduces regulatory, security, and operational risks.

Zero Data Exfiltration

All model inference runs within the client's own infrastructure boundary. No customer data leaves the institution's controlled environment.

Regulatory Alignment

Satisfies data residency requirements under GDPR, UAE PDPL, DIFC Data Protection Law (DPL), ADGM Data Protection Regulations (DPR), and sector-specific guidance.

Deterministic Latency

No dependency on external API rate limits, provider outages, or internet routing. Inference latency is bounded by local compute.

Full Auditability

Every model version, prompt template, retrieval source, and inference output is logged within the institution's audit perimeter.

Architecture Overview

On-Premise Deployment Architecture

Three isolated tiers — ingestion, intelligence, and integration — all executing within the institution's network boundary.

Fig. 1 — High-Level System Architecture
CLIENT INFRASTRUCTURE BOUNDARY DATA INGESTION TIER Core Banking SystemKYC / CDD / EDD Records Transaction MonitorAlerts / STRs / SARs Document StoreIDs / Proof of Address / UBO Sanctions / PEP ListsLocal Mirror — Daily Sync ETL PipelineNormalize — Validate — Chunk Vector Databasepgvector / Milvus (local) AI INTELLIGENCE TIER Local LLM RuntimevLLM / Ollama / TGIGPU: A100 / H100 / L40S RAG EngineRetrieval-Augmented Generation Agent OrchestratorLangGraph / CrewAI (local) KYCAgent EDDAgent TMAgent QCAgent AUDIT LOG — ALL INFERENCE INTEGRATION TIER Case ManagementAuto-populate decisions Analyst DashboardReview / Approve / Override Regulatory ReportinggoAML / FIU Feeds Risk Scoring EngineDynamic risk recalculation REST / gRPC API LayerInternal network only RBAC & mTLSZero-trust auth layer NO EXTERNAL ML API CALLS
Tap diagram to enlarge
Comparison

Cloud ML APIs vs. Local Execution

Traditional approaches rely on sending sensitive data to third-party inference endpoints. Our architecture eliminates this entirely.

Typical Cloud ML Approach
  • Customer PII sent to external inference APIs (OpenAI, Azure, AWS Bedrock)
  • Data may be processed in regions outside regulatory jurisdiction
  • Prompt content potentially used for model training by provider
  • Latency dependent on internet connectivity and API rate limits
  • Vendor lock-in to specific model provider's pricing
  • Audit trail gaps — inference logs held by third party
  • Requires complex Data Processing Agreements
AML Labs — Local Execution
  • All inference runs on institution-controlled GPU infrastructure
  • Data never leaves the designated compliance region
  • No data shared with any ML provider — zero training data leakage
  • Sub-100ms inference latency on local hardware
  • Model-agnostic — swap between Llama, Mistral, Qwen
  • Complete audit trail within institution's SIEM
  • No cross-border data transfer mechanisms needed
Data Flow

Inference Pipeline — Zero External Transfer

A step-by-step view of how a compliance query is processed entirely within the institution's perimeter.

Fig. 2 — Inference Request Lifecycle
1. TRIGGERAlert / Onboardingevent receivedvia internal queue 2. RETRIEVEVector search onlocal embeddingspgvector / Milvus 3. REASONLocal LLM inferencewith RAG contextvLLM on GPU cluster 4. VALIDATEQC agent verifiesoutput + confidencerule-based + LLM check 5. DELIVERStructured result tocase managementREST API — Dashboard PERSISTENT AUDIT LOGEvery stage writes: timestamp, model version, input hash, output, confidence score, retrieval sources, latency NETWORK BOUNDARY — NO OUTBOUND ML TRAFFICFirewall rules block egress to known ML inference endpointsNetwork monitoring alerts on attempted outbound connections — integrated with SOC / SIEM
Tap diagram to enlarge
Technology Stack

Reference Implementation

A proven stack of open-source and enterprise-grade components selected for on-premise deployability and compliance-readiness.

LLM Inference
vLLM / Ollama / TGI
High-throughput local serving with PagedAttention, continuous batching, and quantized model support.
Base Models
Llama 3.x / Mistral / Qwen
Open-weight models fine-tuned on AML/KYC domain data.
Embeddings
BGE / E5 / Nomic Embed
Locally-hosted embedding models for document vectorization.
Vector Store
pgvector / Milvus
On-premise vector database for semantic search over compliance documents.
Orchestration
LangGraph / CrewAI
Multi-agent orchestration with structured compliance workflows.
Document Processing
Unstructured / DocTR
Local OCR and document parsing for IDs and corporate documents.
Compute
NVIDIA A100 / H100 / L40S
On-premise GPU infrastructure, Kubernetes-orchestrated.
Observability
Prometheus / Grafana / ELK
Inference monitoring, drift detection, and audit log aggregation.
Security Architecture

Network Isolation & Trust Zones

Defense-in-depth with distinct network zones enforcing strict ingress/egress rules.

Fig. 3 — Network Security Zones
DMZ Sanctions FeedInbound only — TLS FIU / goAMLOutbound reports — mTLS Reverse ProxyWAF + rate limiting Egress FirewallBLOCK: *.openai.comBLOCK: *.azure.com/openai FW APPLICATION ZONE API GatewaymTLS + JWT + RBAC Analyst UISSO + MFA enforced Agent Orchestrator + RAG EngineNo external network access Local LLM Inference ClusterGPU nodes — no egress permitted Message QueueKafka / RabbitMQ MonitoringPrometheus + Grafana FW DATA ZONE — HIGHEST TRUST Customer PII StoreEncrypted at rest — AES-256 Vector DatabaseEmbeddings — no raw PII Audit DatabaseImmutable append-only log Model RegistryVersioned weights + configs NO EXTERNAL NETWORK ACCESS
Tap diagram to enlarge
Deployment

Implementation Workflow

A structured deployment process ensuring compliance requirements are met from day one.

01

Infrastructure Assessment & Provisioning

Evaluate existing compute infrastructure and provision GPU nodes. Establish network segmentation and firewall rules.

02

Model Selection & Domain Fine-Tuning

Select open-weight base models and fine-tune on anonymized compliance data. All training runs locally.

03

RAG Pipeline & Knowledge Base Construction

Ingest and vectorize institutional knowledge: policies, regulatory guidance, sanctions lists, and case files.

04

Agent Design & Workflow Configuration

Configure specialized agents with structured prompt templates, tool access permissions, and escalation rules.

05

Integration & Validation Testing

Connect to existing systems via internal APIs. Run parallel testing against historical cases.

06

Production Deployment & Monitoring

Gradual rollout with real-time monitoring of inference latency, accuracy, and drift metrics.

Regulatory Compliance

Data Sovereignty Compliance Matrix

Local model execution addresses data handling requirements of major regulatory frameworks.

RegulationRequirementLocal Execution
GDPR (EU)Data minimization, restricted cross-border transfersSatisfied — no external transfer
UAE PDPLPersonal data processed within UAESatisfied — on-premise in UAE
DIFC DPLAdequate data protection for DIFC entitiesSatisfied — local processing
ADGM DPRData protection for ADGM entitiesSatisfied — no third-party sharing
CBUAE AML GuidelinesSecure handling of CDD dataSatisfied — full audit trail
FATF Recommendation 15Controls for new technologies in AML/CFTSatisfied — controlled, auditable AI