Carepath AI

CarePath AI

Inspiration

When someone needs care, the challenge is rarely finding a healthcare facility—it’s knowing which one to trust.

Today, referral decisions are often driven by outdated directories, word-of-mouth recommendations, or whichever facility appears first in a search result. Care planners have little visibility into the quality, evidence, or reliability behind these recommendations.

We built CarePath AI to change that.

Our goal was to create a referral platform that ranks healthcare facilities transparently, explains why each recommendation is made, and openly communicates uncertainty—transforming a blind directory lookup into an evidence-backed decision-making process.

What It Does

CarePath AI helps care planners convert a plain-English patient need into a trusted, ranked shortlist of healthcare facilities.

Key Features

Search and rank 10,000 healthcare facilities from the Virtue Foundation DAIS 2026 dataset based on location, specialty, trustworthiness, and evidence.
AI Copilot (Agent Bricks) that maps natural language requests to specialties, invokes SQL recommendation functions, retrieves evidence, and explains recommendations with trust caveats.
Explain button that generates concise facility summaries using Databricks serving endpoints.
Persistent shortlists using Lakebase for notes, manual rank overrides, and saved facilities.

Each facility receives a composite score balancing trust, evidence quality, and proximity.

score(f) = wt × trust(f) + we × evidence(f) - wd × d(f,pincode)/dmax

This ensures a nearby but poorly verified facility does not automatically outrank a slightly farther, well-evidenced alternative.

How We Built It

Platform: - Databricks Apps - Lakebase - Agent Bricks - Serving Endpoints

Frontend: - React - Vite - TypeScript

Backend: - Node.js - Express - Databricks AppKit

Data Layer: - Medallion-style SQL pipeline (sql/00–03) - workspace.carepath_ai dataset - recommend_by_pincode SQL function

Persistence: - Lakebase Postgres connected through pg.Pool with stable role authentication.

Challenges We Ran Into

Lakebase Integration Mismatch: - AppKit Lakebase plugin targeted Autoscaling deployments while the CLI only supported classic instances. - We connected directly through PostgreSQL and enabled graceful degradation.

Agent Model Selection: - Claude Opus and Gemini were unavailable due to rate limits. - Llama leaked tool calls. - GPT-OSS returned empty responses. - qwen3-next-80b was the only model that consistently produced reliable structured tool calls.

Secret Management: - Removed plaintext credentials from configs. - Migrated secrets into Databricks Secret Scopes.

Token Expiration: - Replaced hourly expiring tokens with a stable service role.

Deployment Workflow: - Improved bundle deployment handling for clean rebuilds and restarts.

Accomplishments We’re Proud Of

Built a complete end-to-end application using 10,000 real healthcare facilities.
Successfully integrated Lakebase, Agent Bricks, and Serving Endpoints.
Built an AI agent that performs real SQL tool calls.
Prioritized transparency with trust caveats.
Implemented secure configuration management and graceful degradation.

What We Learned