Care Convoy

Inspiration

Virtue Foundation's India facility dataset made one thing clear: the hardest part is not finding a hospital. It is deciding whether a medical team can responsibly act on messy, uneven evidence. Care Convoy was inspired by the gap between a data table and a field decision: where should the next specialty care mission go, and what should we verify before moving?

What It Does

Care Convoy is a Track 3 Referral Copilot for India. It helps an operations lead choose a care mission, review the priority district, inspect a lead and backup referral anchor, and decide whether to shortlist, verify first, or hold.

The recommendation is not a black-box score. Care Convoy shows the practical reasons behind each referral plan: local health need, nearby facility coverage, facility fit, trust signals, and cited evidence. When one piece is weak, the app makes that caution visible so the operator knows whether to act now, verify a facility first, or hold the recommendation.

How We Built It

We built Care Convoy as a Streamlit Databricks App using the provided Virtue Foundation tables: facilities, nfhs_5_district_health_indicators, and india_post_pincode_directory.

We orchestrated the referral planner as a gated agent workflow: Need Scout, Supply Mapper, Facility Scout, Trust Verifier, Evidence Auditor, Mission Strategist, and Supervisor each produce a pass/review/block signal, and the Supervisor turns the weakest gate into the final action: shortlist, verify first, or hold.

Unity Catalog and a Databricks SQL Warehouse serve the planning data. Derived Delta tables cache deterministic facility scoring and entity mappings by source-row fingerprint, so changed records can fall back safely instead of reusing stale trust signals. Lakebase persists saved notes and follow-up status. MLflow evaluation checks evidence grounding and operator actionability.

Challenges

The dataset is valuable precisely because it is imperfect. Facility IDs may not represent unique real-world facilities, capability fields can be sparse, district names do not always align cleanly, and public source URLs vary in quality. We turned those risks into product behavior: visible uncertainty labels, citation checks, duplicate flags, and gates that slow the recommendation when evidence is weak.

Accomplishments

Care Convoy goes beyond a map or ranked list. It produces a saveable mission packet with a priority district, lead anchor, backup anchor, confidence, warnings, evidence rows, and a durable follow-up note. The app keeps the workflow readable for an operations lead while still giving judges a clear view of the Databricks architecture behind it.

What We Learned

We learned that data-for-good tools need humility built into the interface. A useful referral tool should not hide uncertainty; it should show which fact needs verification before action. Trust is not one feature; it is a chain across source quality, entity matching, capability evidence, and persistence.

What's Next

Next, we would add travel-time routing, stronger population denominators, Vector Search-backed facility evidence retrieval, and a review queue for analysts who validate weak or duplicate records over time.

Built With

• and-the-provided-virtue-foundation
• databricks-apps
• databricks-asset-bundles
• databricks-model-serving
• databricks-sql-warehouse
• delta-lake
• github
• lakebase
• mlflow-evaluation-and-tracing
• nfhs-5
• pandas
• plotly
• postgresql
• pydeck
• pytest
• python
• streamlit
• unity-catalog