Care India: Verified Healthcare Referral Copilot

An evidence-backed AI referral copilot that helps families and NGOs find trustworthy Indian healthcare facilities while detecting unreliable data and medical deserts.

Comment

Inspiration

In a medical emergency, finding the nearest hospital is not enough. People also need to know whether that hospital can actually provide the required treatment.

Many healthcare records in India are incomplete or inconsistent. A hospital may claim to provide ICU or surgery services, but there may be little evidence about its equipment, staff, oxygen supply, or operation theatres.

We built Care India to help families, NGO workers, and public-health teams find healthcare facilities based on both location and trustworthy evidence.

What it does

A user can enter a normal-language query such as:

Find a hospital near Patna that can handle emergency surgery and has ICU support.

Care India searches more than 10,000 healthcare facility records and shows relevant results on an interactive map.

The system does not blindly trust hospital claims. It checks the available data and shows:

  • Supporting evidence
  • Missing information
  • Trust scores
  • Reasons for recommending or rejecting a facility

It also identifies two different situations:

  • Medical Desert: The required healthcare service is not available in the area.
  • Data Desert: Hospitals exist, but their records are too incomplete to verify.

Key features

  • Natural-language hospital search
  • Interactive map with facility results
  • Evidence-backed recommendations
  • Trust and confidence scoring
  • AI verification of hospital claims
  • Medical Desert and Data Desert detection
  • User reporting for incorrect facility information
  • Clear reasoning for every recommendation

How we built it

The frontend was built using Next.js, React, and React Leaflet. The backend was built with FastAPI.

We used:

  • Databricks Vector Search for finding relevant facilities
  • Unity Catalog and Delta tables for healthcare data
  • LangGraph for the multi-step AI workflow
  • MLflow for tracing and monitoring
  • Codex and GPT-5.6 for planning, coding, debugging, testing, and improving the overall application

How we used Codex and GPT-5.6

We used Codex with GPT-5.6 throughout the development process.

It helped us understand the challenge, plan the system architecture, build FastAPI endpoints, create React components, connect the frontend with the backend, and debug errors.

We also used it to review our code, improve error handling, write test cases, and identify missing production requirements.

GPT-5.6 was especially helpful for reasoning about the complete system. It helped us connect the healthcare problem, data-quality issues, AI verification, backend logic, and user experience.

One important improvement was making sure that mock hospital data could never silently appear as real data. Codex helped us add a clear configuration check and a visible mock-data warning.

Challenges

The biggest challenge was incomplete and contradictory hospital data.

Sometimes the search found many hospitals, but the verification system rejected them because there was not enough evidence about their equipment or services.

At first, we thought the search was broken. Later, we realized this was an important result. This led us to create the Medical Desert and Data Desert feature.

Another challenge was preventing unsupported AI recommendations. We handled this by grounding every recommendation in retrieved facility evidence and clearly showing uncertainty.

What we learned

We learned that trustworthy AI requires more than a powerful model. It also needs good data, evidence, verification, clear limitations, testing, and transparent results.

We also learned that missing data can be an important public-health signal.

Codex and GPT-5.6 helped us build and improve the project much faster by supporting us across planning, development, debugging, testing, and deployment.

Impact

Care India can help:

  • Families find more trustworthy healthcare options
  • NGO workers make safer referrals
  • Public-health teams identify underserved areas
  • Data teams find regions with poor healthcare reporting
  • Users report incorrect or outdated hospital information

What’s next

In the future, we want to add:

  • Live hospital verification
  • Real-time bed and oxygen availability
  • Multilingual and voice search
  • Ambulance routing and travel-time estimates
  • Stronger medical safety checks
  • Better regional healthcare analytics

Built With

Share this project:

Updates