Reference Architecture: Building AI Applications on AWS

AWS as the foundation for enterprise AI

Amazon Web Services offers one of the most complete ecosystems for building production AI applications, from managed foundation models through Amazon Bedrock to the compute, storage, networking, and observability services that surround them. For Australian organisations with existing AWS investments or data residency requirements in ap-southeast-2 (Sydney), a well-designed reference architecture on AWS can accelerate AI Development Australia initiatives while maintaining enterprise security standards.

This article walks through a practical reference architecture TruFyre AI uses when delivering AI Consulting Australia and Custom AI Software projects on AWS.

Architecture overview

• Amazon API Gateway: authenticated API endpoints
• AWS Lambda or Amazon ECS/EKS: application logic and agent workflows
• Amazon Bedrock: managed foundation model access
• Amazon OpenSearch Serverless or Aurora PostgreSQL with pgvector: vector search for RAG
• Amazon S3: document storage and audit archives
• Amazon EventBridge: event-driven indexing pipelines
• Amazon CloudWatch: metrics, logs, and alarms

Amazon Bedrock: model layer

Bedrock provides unified API access to models from Anthropic, Meta, Amazon, and others. Use guardrails for content filtering, implement prompt caching, and log all invocations for audit. Abstract Bedrock behind an internal service for centralised governance.

Compute: Lambda vs ECS vs EKS

Lambda suits request-response AI APIs and lightweight orchestration. ECS on Fargate handles longer agent workflows without cluster management. EKS fits organisations already running Kubernetes; justify the operational complexity with platform team capacity.

Data layer: OpenSearch vs Aurora pgvector

OpenSearch Serverless excels at hybrid search and large document corpora. Aurora PostgreSQL with pgvector suits teams wanting vectors alongside relational application state. Many architectures use both.

Event-driven patterns with EventBridge

S3 upload triggers EventBridge, Lambda extracts and embeds documents via Bedrock, vectors write to OpenSearch or pgvector, and completion events notify downstream systems.

Security and networking

• Private subnets with VPC endpoints for Bedrock, S3, and Secrets Manager
• IAM roles: never long-lived access keys in code
• Encryption at rest and in transit throughout
• AWS WAF on public API Gateway endpoints
• CloudTrail for API audit logging

Observability with CloudWatch

Track token usage, latency percentiles, retrieval hit rates, and cost anomalies. Structured logs with correlation IDs link user requests to model calls.

Cost management

Tag resources by application and environment. Use Cost Explorer budgets with alerts. Right-size OpenSearch and apply S3 Intelligent-Tiering for archives.

Key takeaways

• AWS provides end-to-end building blocks for Enterprise AI production workloads.
• Choose compute and vector storage based on workload patterns.
• Event-driven ingestion keeps APIs responsive at scale.
• Security, observability, and cost tagging are non-negotiable.

Building AI applications on AWS with a deliberate reference architecture lets Australian enterprises leverage existing cloud investments while meeting the operational bar that AI Integration at scale demands.