🏗️ Architecture Options

Two paths to build the AML Investigator Copilot

🔧

Option A

Modular Microservices

76%

Team Size 9 people

MVP 4 months

v1.0 8 months

3-Year TCO $5.85M

Best For Enterprise scale

🕸️

Option B

Graph-Centric RAG

79% ⭐

Team Size 5 people

MVP 2.5 months

v1.0 5 months

3-Year TCO $2.97M

Best For Fast iteration

📊 Detailed Ratings

Criterion	Option A	Option B	Winner
Market Fit	85%	80%	🔧 A
Technical Feasibility	75%	85%	🕸️ B
Time to Market	60%	90%	🕸️ B
Scalability	90%	70%	🔧 A
Cost Efficiency	65%	90%	🕸️ B
Competitive Moat	80%	60%	🔧 A

🔧 Option A: Modular Microservices

Philosophy: Specialized models per task, maximum customization, enterprise-grade scalability.

Architecture

┌─────────────────────────────────────────────────────────────┐ │ INVESTIGATOR UI (React) │ └─────────────────────────────────────────────────────────────┘ │ API GATEWAY (Kong) │ ┌─────────────────────┼─────────────────────┐ ▼ ▼ ▼ ┌───────────────┐ ┌───────────────┐ ┌───────────────┐ │ ENTITY SERVICE│ │DOCUMENT SERVICE│ │REASONING SERVICE│ │ • Resolution │ │ • OCR/Extract │ │ • Inconsistency│ │ • Risk Score │ │ • Classification│ │ • Capacity │ │ • Network Map │ │ • NER/Parsing │ │ • Signals │ └───────────────┘ └───────────────┘ └───────────────┘ │ │ │ └─────────────────────┼─────────────────────┘ │ MESSAGE BUS (Kafka) + EVENT SOURCING │ ┌─────────────────────┼─────────────────────┐ ▼ ▼ ▼ PostgreSQL Elasticsearch Milvus (Cases/Audit) (Full-text) (Vectors) │ MODEL SERVING (vLLM) 4-8x A100 GPUs

Tech Stack

LLMLlama 3.1 70B

ServingvLLM + Triton

GraphNeo4j

VectorMilvus

SearchElasticsearch

QueueKafka

DBPostgreSQL

K8sOpenShift

Timeline

Months 1-2

Infrastructure + core data pipeline

Months 2-3

Document processing + basic entity service

Months 3-4

Simple reasoning (8B model) + basic UI → MVP

Months 5-8

Fine-tuned models, full UI, audit system → v1.0

Cost (3-Year)

$1.25M

Infrastructure

$4.2M

Personnel (9 FTE)

Pros

Maximum customization & flexibility
Fine-tuned models = higher accuracy
Scales to Tier-1 bank volumes
Stronger competitive moat

Cons

8 months to v1.0 — competitors moving fast
$5.85M 3-year cost
Requires 9-person team
Higher integration complexity

🕸️ Option B: Graph-Centric RAG

Philosophy: Unified knowledge graph + single LLM, simpler operations, faster iteration.

Architecture

┌─────────────────────────────────────────────────────────────┐ │ INVESTIGATOR UI (Next.js) │ └─────────────────────────────────────────────────────────────┘ │ UNIFIED API (FastAPI) │ ┌─────────────────────────────────────────────────────────────┐ │ ORCHESTRATION (LangGraph) │ │ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ │ │ │ Pre-Decision│ │ Document │ │ Ad-hoc │ │ │ │ Agent │ │ Agent │ │ Query │ │ │ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ │ │ └────────────────┼────────────────┘ │ │ ▼ │ │ UNIFIED LLM (Llama 70B / Mixtral) │ └─────────────────────────────────────────────────────────────┘ │ ┌─────────────────────┼─────────────────────┐ ▼ ▼ ▼ KNOWLEDGE DOCUMENT STORE AUDIT LOG GRAPH PostgreSQL + PostgreSQL (Neo4j) pgvector (Event Source)

Tech Stack

LLMLlama 70B / Mixtral

OrchestrationLangGraph

GraphNeo4j

Vectorpgvector

DocsUnstructured.io

QueueRedis Streams

DBPostgreSQL

K8sK3s

Timeline

Weeks 1-3

Infrastructure + graph setup + ingestion

Weeks 4-6

LLM integration, basic agents, copilot UI

Weeks 7-10

Document extraction, entity 360 → MVP

Months 3-5

Advanced graph, audit system, export → v1.0

Cost (3-Year)

$494K

Infrastructure

$2.33M

Personnel (5 FTE)

Pros

2.5 months to MVP — fast market validation
Half the cost ($2.97M vs $5.85M)
5-person team — leaner operations
Faster iteration on prompts

Cons

May hit scale limits at Tier-1 banks
Less model customization
Easier for competitors to replicate
Prompt engineering can be fragile

💡 Recommendation

Start with Option B, evolve toward Option A

Get to market in 2.5 months with core functionality. Learn from real users. Add Option A components (fine-tuned models, specialized services) only where needed. If traction with large banks in Year 2, invest in full microservices architecture.

🎯 Progressive Complexity Approach

Phase 1: MVP (Option B)

Get to market in 2.5 months with core copilot functionality

Phase 2: Learn

Understand actual workflows and pain points from real users

Phase 3: Scale Selectively

Add Option A components only where data shows you need them

Phase 4: Enterprise (Year 2+)

If traction with large banks, invest in full Option A infrastructure

        This optimizes for:
        Fast market validation
Lower initial capital burn
Reduced technical risk
Flexibility to pivot based on learnings

      