Veterinary medicine is high-stakes, low-margin and chronically under-staffed. We built a multi-agent diagnostic assistant to help clinicians triage cases faster without compromising clinical judgement.
The Orchestrator Pattern
A single supervisor agent owns the conversation and routes sub-tasks to specialist agents: history-taking, differential generation, RAG-grounded reference lookup, and a safety reviewer.
RAG-Grounded Medical Reasoning
Every diagnostic claim must cite a source from our curated veterinary corpus. Uncited claims are blocked at the safety layer.
Safety Guardrails
A dedicated reviewer agent runs after every response. Latency cost around 400ms. Worth every millisecond.
What We Would Revisit
We would build the eval harness before the agents. We would use LangGraph from day one. And we would invest earlier in structured output validation with Pydantic.
