Project 02
Multi-Agent Research Assistant
A LangGraph supervisor routing to specialist agents, exposed over MCP
A supervisor agent decomposes a research question and routes to specialists: a retriever that queries the knowledge base, a summariser, a fact-checker that re-grounds every claim, and a SQL analyst for structured questions. State, retries and an approval gate are explicit graph nodes rather than prompt instructions. The same tools are published as an MCP server, so Claude Desktop or any MCP client can query the knowledge base directly.
Try it
How it works
- 01Plan
Supervisor node decomposes the question into a task list held in graph state.
- 02Route
Conditional edges dispatch each task to the specialist that can answer it.
- 03Verify
Fact-checker re-retrieves evidence for each claim; unsupported claims are dropped, not softened.
- 04Approve
Interrupt node pauses the graph for human approval before any write action.
- 05Expose
The same tool definitions served over MCP for external clients.
Infrastructure & running cost
| Service | Role | Cost |
|---|---|---|
| Lambda (container) | Graph execution, same image as project 1 | Free tier |
| DynamoDB | Graph checkpointer — survives Lambda cold starts | Free tier |
| Lambda Function URL | Streaming SSE responses, no API Gateway needed | $0 |
Closes the largest gap between the CV and current agent-engineering job specs.