Home
Multi agent maritime dark vessel investigation on the AWS generative AI stack: Strands Agents and LangGraph agents on Amazon Bedrock AgentCore, four MCP tool servers over PostGIS behind a policy enforcing gateway, a FastAPI watch floor where officers review every AI draft, OpenTelemetry into Grafana and CloudWatch, and a CDK deployment for all of it.
| If you want to | Read |
|---|---|
| Run it on your laptop in one command | Getting Started |
| Understand the system end to end | Architecture |
| See how the agents, models and prompts fit together | Agents and Models |
| Know what tools the agents can call and how calls are policed | Tool Plane |
| Learn the tables and the provenance guarantees | Data Model |
| Use the watch floor as an officer | Watch Floor |
| Operate it on AWS: deploy, pause, alarms, rotation | Operations |
| Check what is measured and what gates a change | Evaluations |
| Review the security posture | Security |
| Read why a design choice was made | Decisions |
| Fix something that broke | Troubleshooting |
Alert (a detector finding the Watch agent raised, an AI draft until reviewed), Sweep (one pass of the detectors and the Watch agent over the last twelve hours), Investigation (one run of the orchestrator graph for one vessel), VOI Report (the Vessel of Interest report an investigation produces), Action (a tasking proposal that waits for approval). Review state (draft, accepted, rejected) is separate from approval; see CONTEXT.md.
This wiki summarises and links. The repository documents are the source of truth: README, ARCHITECTURE.md, and the docs set (technical reference, API, use cases, runbook, security, audit, ADRs).