CareReach

Databricks-native AI agent that uses trust-weighted evidence to distinguish real healthcare gaps from data gaps across India’s districts.

Comment

Inspiration

Healthcare planners in India face a critical and costly challenge: when a district shows a low facility count, does that mean there's a real healthcare gap requiring urgent intervention — or is it simply missing/unreported data? Getting this wrong wastes millions of dollars and costs lives. We built CareReach to solve this using trust-weighted evidence aggregation, so planners can confidently distinguish real gaps from phantom ones.

What it does

CareReach is an AI-powered healthcare access planning system that:

  • Ingests 10,000+ government healthcare facility records, 165,627 PIN codes, and 109 NFHS-5 district health indicators across 707 districts in India
  • Uses Databricks Foundation Models (DBRX Instruct) to AI-enrich facility descriptions, extracting clinical readiness, accessibility barriers, quality trajectory, and workforce stability
  • Computes trust scores (0–1) for every facility, calibrated by the AI's own data-reliability assessment
  • Characterizes each district as REAL_GAP, DATA_GAP, MIXED_SIGNALS, or ADEQUATE_COVERAGE
  • Prioritizes interventions as P1_IMMEDIATE → P4_MONITOR based on composite gap scores
  • Serves all of this through an MLflow AI planning agent you can query in natural language
  • Presents everything in a modern Gradio web app deployed as a Databricks App, featuring a real-time KPI dashboard, geospatial choropleth maps, analytics, and a live AI chat interface

How we built it

We built CareReach entirely on the Databricks platform using a full medallion architecture:

  • Bronze Layer: Raw ingestion of healthcare facilities, PIN code directory, and NFHS-5 health indicators into Delta Lake (Unity Catalog)
  • Silver Layer: Geographic enrichment via PIN code joins, AI-powered description analysis using the Databricks Foundation Model API, trust score computation, deduplication, and field coverage analysis
  • Gold Layer: District and state-level aggregated analytics tables (gold_care_gap_analysis, gold_facility_trust_aggregates, gold_demand_supply_overlay)
  • AI Agent: Built with the Databricks Agent Framework and MLflow, the agent queries gold-layer tables via three custom tools — gap analysis, facility search, and demand-supply analysis
  • App: A Gradio interface deployed as a Databricks App, with Plotly and Folium for interactive visualizations

Challenges we ran into

  • Data quality ambiguity: Designing a trust scoring framework that is both statistically meaningful and interpretable to non-technical planners was iterative
  • AI calibration: Getting the Foundation Model to reliably self-assess its own confidence in the data it was analyzing required careful prompt engineering
  • Scale: Processing 10,000+ facility records with AI enrichment at the silver layer required efficient batching and Serverless SQL Warehouse optimization
  • Gap classification thresholds: Determining the right trust score and facility count thresholds to distinguish REAL_GAP from DATA_GAP required domain reasoning and testing

Accomplishments that we're proud of

  • Identified 200+ phantom facility regions (DATA_GAPs) that would have triggered wasteful investigations under traditional methods
  • Flagged 150+ districts for P1_IMMEDIATE or P2_HIGH intervention with high confidence
  • Achieved 85%+ gap characterization confidence for priority regions
  • Built a fully end-to-end Databricks-native solution: Unity Catalog → Delta Lake → Foundation Models → MLflow Agent → Databricks App
  • Integrated 3 distinct government datasets (facilities, PIN codes, NFHS-5) into a unified analytics layer

What we learned

  • Trust-weighted evidence aggregation is a powerful pattern that goes beyond healthcare — it's applicable anywhere data completeness is uneven
  • The Databricks Agent Framework makes it surprisingly straightforward to wire up natural language queries to structured data via custom tools
  • Gradio + Databricks Apps is a fast and effective combination for deploying data-heavy interactive applications
  • AI enrichment of semi-structured text (facility descriptions) can surface signals that are impossible to capture with structured data alone

What's next for CareReach

  • Real-time ingestion with Delta Live Tables for streaming facility updates
  • Predictive gap forecasting — anticipating emerging gaps before they become crises
  • Mobile app for field data collection and GPS-based facility verification in remote areas
  • Multi-country expansion — the medallion pipeline is modular and can adapt to other developing nations
  • Integration with national health ministry APIs for automated reporting and real-time alerts
  • Computer vision for facility image analysis and sentiment analysis of patient reviews

Built With

Share this project:

Updates

Submission history