Autonomous Multi-Agent Workflow Orchestrator

Designed and shipped a production multi-agent orchestration system in Python, where five specialised Claude agents collaborate via typed JSON contracts (pydantic-validated) to complete complex business reporting tasks end-to-end, with retry-and-fallback routing so the pipeline degrades gracefully rather than failing silently.

Orchestrated tool use across agents using the Claude API: web search, PDF ingestion, structured data extraction, and a citation tracker that registers every source and rejects any finding referencing an unregistered citation id. Hallucinated sources stopped being a class of bug.

Built a real-time Streamlit execution dashboard that streams every inter-agent JSON message as a colour-coded card, so non-technical operators can see exactly which agent is on the critical path of a run. Persistent ChromaDB memory per stakeholder lets the researcher skip work already verified in previous runs.