knightingale-agent

The Knightingale automated nurse-scheduling voice agent.

This is the long-running brain of the system: a LiveKit agent that answers/forwards calls via Twilio, talks using OpenAI Realtime, and decides which nurse to call by reading the decision policy in skills/ and querying Supabase.

What lives where

• This repo (self-hosted, long-running): the LiveKit voice agent + decision policy.
• claude repo (Vercel, serverless): stateless HTTP endpoints — SMS sending, webhooks.
• Supabase: data (facilities, nurses, approvals, availability, shifts), the get_candidate_pool function, and the learned_decisions store.

Why the skill lives here

skills/nurse-selection.md is the policy the agent reads before every selection. It lives in Git — not the database — so every change is a reviewable, version-controlled commit. The agent reads this file; it never rewrites it. Learning happens in the Supabase learned_decisions store (records of Paul's actual rulings), never by editing policy. That separation is what keeps the system auditable.

Structure

knightingale-agent/
  skills/
    nurse-selection.md      # the decision policy (read by the agent, edited by Paul)
  agent/                    # (Step 11) the LiveKit + OpenAI Realtime agent code
  README.md

Status

• Supabase data layer + get_candidate_pool function
• Decision skill file (v0.2)
• Reliability scoring (0–100, recency-weighted, from call_events)
• learned_decisions store
• LiveKit + OpenAI Realtime agent