Virtue Health Explorer

Virtue Health Explorer turns India's raw healthcare data into clear answers, showing where care is missing and helping NGOs, officials, and patients decide where to send it next.

Comment

Inspiration

Behind every row in a healthcare dataset is a person who can't reach a doctor. In India, a mother in a district with no nearby maternity facility, a patient who needs dialysis three hours away, a child missing routine immunisation — these aren't statistics, they're lives shaped by where a hospital was or wasn't built. The data to find these people exists: the DAIS 2026 Virtue Foundation dataset holds ~10,000 facilities and NFHS-5 health indicators for 706 districts. But it sits raw and unreadable to the very people who could act on it — the NGO planning its next medical mission, the official deciding where to send resources, the family searching for care tonight. We built Virtue Health Explorer to close that gap: to turn data that describes suffering into a tool that helps reduce it.

What it does

Virtue Health Explorer finds where care is missing, explains why, and points to what would help — for three people who each need a different answer.

  • It pinpoints the places that need help most. A Care Access Gap score (0–100) for every state, built from Availability, Accessibility, Specialty coverage, and Affordability and measured against WHO benchmarks (3.0 beds / 1.0 doctors per 1,000), so limited resources can go where they'll change the most lives.
  • It explains the human story behind the score. Expand any state and you see the real outcomes underneath — immunisation, institutional births, anaemia, child stunting, insurance, sanitation — the lived realities the numbers stand for.
  • It surfaces the hardest-hit districts — the "medical deserts" where coverage is worst — and the specialties most desperately scarce, so a mission can be aimed precisely.
  • It answers anyone's question in plain English. A grounded AI assistant lets a planner, an official, or a patient ask what they actually want to know, and answers from real data — showing exactly which sources it used and being honest about what the data can't promise. For an NGO it drafts a mission plan; for a family it can find the nearest facility that offers the care they need.
  • It helps people find care by describing it — "trauma-capable hospitals near rural Odisha," "a centre offering dialysis" — and semantic search finds the closest match.

Every persona — the mission planner, the public-health official, the patient — gets a version of the tool shaped around the decision in front of them.

How we built it

We built it entirely on Databricks, with a medallion architecture so that every insight is trustworthy enough to act on:

  • Source → Silver → Gold (Unity Catalog). A read-only Delta Share feeds a clean silver layer (we geo-correct facility locations so a hospital is never misplaced into the wrong state), then we precompute every gap, desert, scarcity score, and recommendation into a gold layer in SQL. The app does zero runtime math — the analysis is reproducible, not guessed.
  • Lakebase (serverless Postgres) syncs the gold tables for instant reads, stores saved mission plans, and powers semantic search via pgvector.
  • The app — React + Vite + TypeScript on Express, built with Databricks AppKit.
  • The AI assistant is an owned tool-calling agent whose nine tools reuse the exact same SQL as the rest of the app, so a number it tells a planner is the same number on the dashboard. Facility embeddings come from bge-large-en.

Challenges we ran into

  • Earning trust with imperfect data. Bed and doctor counts are genuinely missing for most facilities, and ~82 source records were corrupt. We refused to paper over it: bad rows are quarantined and never served, and the assistant discloses the data's limits on every affected number, using them only for relative comparison. When a tool helps decide where care goes, honesty isn't optional.
  • Keeping the AI grounded. We forced every quantitative claim through a deterministic tool and made the assistant show its sources — so it informs decisions instead of inventing them.
  • Making it real and fast — sub-10ms reads via Lakebase, a shared-pipeline workaround for a strict quota, and resilient batch-embedding (rate-limit backoff, dedup, resumable) so search can eventually cover every facility.

What we learned

The hardest part of a tool meant to help people isn't the model — it's trust. Precomputing analytics deterministically, grounding every answer in real data, showing provenance, and being relentlessly honest about gaps is what turns a dashboard into something a person can actually act on with confidence. We also saw how directly the right data architecture can shorten the distance between a raw dataset and a decision that puts care where it's needed.

What's next

Right now Virtue Health Explorer shows where care is missing. Our next chapter is helping close those gaps — turning insight into action that reaches people.

  • From "where the gap is" to "what to do about it." We want every gap to come with a costed, on-the-ground plan: how many beds, which specialists, what a mission would take to move a district out of the "desert" zone — so an NGO can act in days, not quarters.
  • Connecting the people with the need to the people who can help. Pairing under-served districts with the foundations, mobile clinics, and volunteer specialists best positioned to serve them — using the tool as a matchmaker, not just a map.
  • Putting it in the hands of the families who need care. A lightweight, low-bandwidth, multilingual version a patient or community health worker can use on a basic phone to find the nearest facility that actually offers the care they need.
  • Measuring the change we set out to make. As newer data arrives, tracking whether gaps actually shrink where missions are sent — so impact is proven, not assumed, and resources keep flowing to where they help most.
  • Helping the helpers see ahead. Forecasting where access will erode next as populations grow, so care can be planned before a community falls behind instead of after. Our north star is simple: shorten the distance — in kilometres, in rupees, and in time — between a person who needs care and the moment they receive it.

Built With

Share this project:

Updates

Submission history