Inspiration

Healthcare referrals are often slowed down by fragmented facility information, uncertain capabilities, and manual verification. We wanted to build a copilot that helps coordinators quickly find defensible referral options while being honest about what the data can and cannot prove.

What it does

Referral Copilot guides a coordinator through a referral request, searches governed facility data, ranks candidate facilities, and generates an evidence-grounded handoff brief. It highlights capability match, contact confidence, location fit, and record quality, while reminding users to verify availability and acceptance directly.

How we built it

We built the app on Databricks Apps with AppKit and Agent Bricks, using a medallion architecture:

Silver table: medallion_architecture.silver.silver_facilities
Gold tables: contact, location, procedure, and specialty facts
Lakebase: stores referral cases and coordinator decisions
Databricks SQL: powers repeatable analytics queries
Agent Bricks: provides an Agent Bricks Supervisor endpoint for referral reasoning
Model Serving/AppKit: connects the app to the governed agent endpoint

Challenges

The hardest parts were normalizing messy facility data, avoiding unsupported claims, and making the app resilient when SQL fields arrived as arrays, objects, or strings. We also had to shift from a single indexed gold table to a cleaner silver/gold table design without breaking the user workflow.

What we learned

We learned how to combine Databricks Apps, Lakebase, SQL Warehouses, Agent Bricks, and medallion tables into one governed application. Most importantly, we learned that in healthcare workflows, the best AI experience is not just confident. It must be traceable, cautious, and clear about uncertainty.