Architecture patterns for enterprise Agentic AI across GCP, AWS Bedrock, and Azure OpenAI.
Built from real production deployments — not theoretical diagrams.
Status: Production deployed · Cloud Run asia-south1 · June 2026
Reference implementation: retail-agentic-ai-platform
┌─────────────────────────────────────────────────────────┐
│ CLIENT APP │
│ Web · Mobile · Chat · In-store │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ CLOUD LOAD BALANCING + CLOUD ARMOR │
│ WAF · DDoS Protection · SSL Termination │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ API GATEWAY │
│ JWT Auth · Rate Limiting · Routing │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ CLOUD RUN — LANGGRAPH CONTROL PLANE │
│ │
│ Governor Agent (keyword-first routing) │
│ ├── Search Executor → Vertex AI Search │
│ ├── Transaction Executor → Risk Engine │
│ ├── RAG Agent → BigQuery policy_chunks │
│ └── Chat Node → Gemini direct │
│ ↓ │
│ Generator (Gemini 2.5 Flash) │
│ ↓ │
│ Policy/Watcher (compliance gate) │
└──────────────────────┬──────────────────────────────────┘
↓
┌──────────────┬───────────────┬──────────────────────────┐
│ Vertex AI │ Gemini 2.5 │ Model Armor │
│ Search │ Flash │ Content Safety │
└──────────────┴───────────────┴──────────────────────────┘
↓
┌──────────────┬───────────────┬──────────────────────────┐
│ BigQuery │ Firestore │ Cloud DLP │
│ Analytics │ State/HITL │ PII Scanning │
└──────────────┴───────────────┴──────────────────────────┘
Key design decisions:
| Decision | Rationale |
|---|---|
| LangGraph over custom orchestration | State machine with conditional edges — easier to debug and extend |
| Keyword-first Governor | Zero LLM cost on 80% of queries — only ambiguous queries hit Gemini |
| Vertex AI Search over custom search | Managed semantic search with personalization, no embedding pipeline needed |
| Cloud DLP on all outputs | PII leakage prevention at the infrastructure layer, not the prompt layer |
| Firestore for HITL queue | Real-time updates, TTL for GDPR, serverless scaling |
Status: Production delivered · AWS Lambda + API Gateway · Wipro (2026)
┌─────────────────────────────────────────────────────────┐
│ CLIENT APP │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ API GATEWAY (REST) │
│ JWT Auth · Rate Limiting · WAF │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ AWS LAMBDA — ORCHESTRATOR │
│ │
│ Extraction Agent (Claude 3.5 Sonnet) │
│ → Parses unstructured documents │
│ → Extracts structured entities │
│ │
│ Inference Agent (Claude 3.5 Sonnet) │
│ → Applies business rules │
│ → Generates compliance assessment │
│ │
│ HITL Validator │
│ → Confidence < threshold → human review │
│ → Audit trail → DynamoDB │
└──────────────────────┬──────────────────────────────────┘
↓
┌──────────────┬───────────────┬──────────────────────────┐
│ Bedrock │ DynamoDB │ S3 │
│ Claude 3.5 │ Audit Trail │ Document Store │
└──────────────┴───────────────┴──────────────────────────┘
Key design decisions:
| Decision | Rationale |
|---|---|
| Two specialized agents over one | Separation of concerns — extraction errors don't corrupt inference |
| Lambda over containers | Compliance workloads are bursty — serverless scales to zero when idle |
| DynamoDB for audit | Sub-millisecond writes, TTL for data retention, no schema lock-in |
| Confidence-based HITL | Not all decisions need human review — only those below the confidence threshold |
Outcomes achieved:
- 60% reduction in manual compliance processing
- 100% audit traceability — zero regulatory findings
Status: Production delivered · AKS + Azure ML · Ericsson (2022-2024)
┌─────────────────────────────────────────────────────────┐
│ CLIENT APP │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ AZURE API MANAGEMENT │
│ Auth · Rate Limiting · Monitoring │
└──────────────────────┬──────────────────────────────────┘
↓
┌─────────────────────────────────────────────────────────┐
│ AKS — ML INFERENCE CLUSTER │
│ │
│ RAG Pipeline │
│ → Pinecone vector retrieval │
│ → Azure OpenAI embedding │
│ → GPT-4 grounded generation │
│ │
│ MLOps Layer │
│ → MLflow experiment tracking │
│ → Drift monitoring │
│ → Automated retraining triggers │
└──────────────────────┬──────────────────────────────────┘
↓
┌──────────────┬───────────────┬──────────────────────────┐
│ Azure │ Databricks │ Azure Monitor │
│ OpenAI │ Delta Lake │ Telemetry │
└──────────────┴───────────────┴──────────────────────────┘
Key design decisions:
| Decision | Rationale |
|---|---|
| AKS over Azure Functions | ML workloads need persistent GPU resources — serverless cold starts unacceptable |
| Pinecone for vector storage | At 100k+ documents, managed vector DB outperforms in-memory FAISS |
| MLflow for experiment tracking | Open-source, portable across clouds, integrates with Databricks |
| Databricks Delta Lake | ACID transactions on ML feature data — critical for audit traceability |
Outcomes achieved:
- 50% faster KYC processing
- 12% fraud detection accuracy improvement
- 15% cloud cost optimisation via serverless design
| Capability | GCP | AWS | Azure |
|---|---|---|---|
| Best LLM | Gemini 2.5 Flash | Claude 3.5 Sonnet | GPT-4o |
| Managed Search | Vertex AI Search | Kendra | Azure Cognitive Search |
| Vector DB | BigQuery Vector Search | OpenSearch | Azure AI Search |
| Serverless | Cloud Run | Lambda | Azure Functions |
| Orchestration | LangGraph on Cloud Run | Bedrock Agents | Azure AI Studio |
| Data Platform | BigQuery | Redshift | Synapse Analytics |
| Compliance | Cloud DLP + Model Armor | Macie + Guardrails | Azure Content Safety |
| Best for | GCP-native, retail/search | Enterprise, regulated | Microsoft shop, FinTech |
Choose GCP when:
→ You need Vertex AI Search for Commerce
→ BigQuery is your data platform
→ LangGraph + Gemini is the target stack
→ GCP-native data sovereignty required
Choose AWS when:
→ Claude 3.5 Sonnet is the required model
→ Existing AWS infrastructure
→ Lambda/serverless-first architecture
→ Strong Bedrock Agents ecosystem preferred
Choose Azure when:
→ Microsoft enterprise agreements in place
→ GPT-4 with Azure compliance controls needed
→ Teams/Office 365 integration required
→ European data residency with GDPR focus
Nadeem Ahmad — Principal Enterprise AI Architect
15+ years across GCP, AWS, and Azure in telecom, fintech, and retail
📧 nadeem.ahmad.arch@gmail.com
🔗 LinkedIn
📅 Book a session