09 / AI ENGINEERING / CASE STUDY
NovaMart Multi-Agent RAG
AWS agent orchestration connecting policy retrieval with business-workflow tools.
Private implementation Private source
Architecture
Agent router → Policy retrieval → Workflow tools → Trace / log hooks
Problem & context
A support workflow needs to combine policy knowledge with order/inventory actions while tracking the request across tools.
Objectives
Explore specialized agents, retrieval-backed policy context, workflow state and observability on AWS.
My role & project work
Developed Python multi-agent orchestration with Bedrock/Strands integrations, Knowledge Base retrieval helpers and DynamoDB/Lambda workflow tools.
Technical implementation
Authorized private-source review confirms orchestration, policy retrieval, workflow-state helpers, Guardrail application hooks, AgentCore deployment tooling and CloudWatch/X-Ray observability code. This public summary intentionally contains only high-level capabilities.
Engineering decisions
Separate policy retrieval, workflow actions and response composition into specialized agent/tool responsibilities. Keep observability code alongside orchestration so requests can be traced across boundaries.
Challenges & solutions
Working integration code still needs runtime credentials, configured cloud resources and end-to-end tests. The presence of deployment tooling does not prove a current deployed AgentCore runtime.
Testing & verification
Reviewed the owner-authorized private repository's current source structure and implementation symbols. No private source, configuration, resource identifiers or screenshots are republished. Cloud resources, model calls and tests were not executed.
Current status
Private source implementation. A high-level walkthrough can be shared on request; no current cloud-runtime or benchmark result is claimed.
Lessons & next steps
A multi-agent project is stronger when retrieval, state transitions, tool authorization and request traces can be demonstrated independently.