Asclepius

Inspiration

The Virtue Foundation gave us ~10,000 real Indian healthcare-facility records — and a hard truth: the data is messy, self-reported, and unverified. In a country where 143M people in low- and middle-income regions wait for surgery each year, planners, patients, and hospitals are all making life-affecting decisions on top of data nobody can fully trust.

The four hackathon tracks each attack a different symptom of that same problem. We didn't want to pick one — because the same untrusted record is what fails a patient looking for care, a planner mapping a medical desert, and a hospital deciding who to hire. So we built one platform that answers all four, with a single rule: never present a weak claim as fact.

What it does

Asclepius is a three-sided healthcare-intelligence platform on Databricks:

• Facility Trust Desk (Track 1) — opens any facility's claimed capabilities, cites the underlying record text, and is honest about what's verified vs. self-reported.
• Medical Desert Planner (Track 2) — ranks 706 districts by distance-based, per-capita care scarcity and, crucially, separates a real care desert from a merely data-poor district (a join gap is never scored as scarcity).
• Referral Copilot (Track 3) — a patient enters a location and a need and gets an evidence-attached shortlist; a grounded AI assistant answers in plain language (and Hindi), citing verbatim facility text.
• Data Readiness Desk (Track 4) — a human-in-the-loop review queue that surfaces contradictions, suspicious claims, sparse fields, and high-leverage records to fix before the data is trusted for planning.
• Plus a doctor↔hospital marketplace, an India coverage Atlas (care coverage × NFHS-5 health-condition overlay), and full persistence of every note, review, shortlist, and decision.

Every score, ranking, and recommendation ties back to source text and carries an honest confidence level.

How we built it

• Databricks Apps (Free Edition) — one React/TypeScript + Node app (AppKit), four personas, one product.
• Lakebase (Postgres) — all OLTP persistence (accounts, reviews, referrals, review actions) plus synced read tables; a medallion in Unity Catalog (raw → silver → gold) feeds it.
• Databricks Foundation Models (Llama 3.3 70B) power the assistant, behind a literal-substring citation guard that re-anchors every quote to the real source bytes and refuses anything it can't ground.
• Mosaic AI Vector Search + SQL warehouse for retrieval and analytics.
• A rigorous distance-based desert model over facility claims, NFHS-5 need, and Census-2011 population; stable views make any future data re-clean a one-line re-point instead of a rebuild.

Challenges we ran into

• Free Edition constraints shaped everything: a 3-app cap, 24h auto-stop, one Vector Search endpoint, one sync pipeline, and daily compute limits — we even made the build warehouse-independent when the SQL warehouse got disabled mid-event.
• The data doesn't contain doctors, symptoms, or outcomes — so we bridged honestly: NFHS-5 prevalence as a demand proxy, an LLM symptom→specialty map, and a user-seeded marketplace, all clearly labeled as modeled.
• Telling a real gap from a data gap was the hardest problem; the naive gold score pinned every unmapped district to maximum scarcity, so we rebuilt it to flag 189 "unknown-supply" districts separately.

What we learned

Communicating uncertainty is a feature, not a disclaimer — and architecting for swap-safety (stable views over a medallion) let the data team keep re-cleaning while the app kept running. The multi-tool Databricks stack (Lakebase + Foundation Models + Vector Search + Unity Catalog) composed far more cleanly than we expected.

What's next

Live NMC/IMR doctor-registration verification (a "✓ verified" badge), per-capability trust grading wired from our corroboration engine, a district-level choropleth, and nationwide patient geography.

Built With

• databricks
• databricks-apps
• databricks-foundation-models
• databricks-sql
• lakebase
• llama-3.3-70b
• maplibre-gl
• mosaic-ai-vector-search
• nfhs-5
• node.js
• phosphor-icons
• playwright
• postgresql
• python
• react
• recharts
• tanstack-query
• typescript
• unity-catalog
• vite