Natnael Mengesha

Enterprise GenAI on Bedrock

One governed Bedrock plane: standardized FM access, RAG, guardrails, private data paths.

Amazon Bedrock · RAG · Guardrails · Step Functions · Lake Formation

1
governed FM entry point
100%
model calls audited
0
direct internet egress
CALLERS & ORCHESTRATION Enterprise apps internal clients Event triggers Amazon EventBridge Orchestration AWS Step Functions BEDROCK PLANE Guardrails safety + PII filters Foundation models Amazon Bedrock Agents Bedrock Agents + actions PRIVATE DATA PLANE Governance Lake Formation + KMS Ingestion Amazon S3 + AWS Glue Knowledge Bases Amazon Bedrock Vector stores OpenSearch / pgvector prompt filtered input trigger invoke model calls retrieve vector search data sync access policy
Fig. 1 — Guardrailed Bedrock plane over a governed private data plane. No direct internet egress.

solid = account · long dash = VPC · short dash = zone or planned

Context

  • Enterprise teams wanted GenAI. No standard path to foundation models.
  • Sensitive data in scope. PII risk. No direct internet egress permitted.
  • Ad hoc prompts. No evaluation baseline. No shared orchestration.

Decision

  • One Bedrock plane. Standardized FM access behind Guardrails for safety and PII.
  • RAG via Knowledge Bases. OpenSearch Serverless and Aurora pgvector stores.
  • Agents call internal APIs. Step Functions and EventBridge orchestrate.
  • S3 + Glue ingestion under Lake Formation. KMS at rest. Private paths only.

Outcome

  • Standardized FM access across the enterprise. CloudTrail audits every call.
  • Prompts versioned, evaluated against golden examples, shipped through CI/CD.
  • Zero direct internet egress on GenAI data paths.