Quantum-Health

QuantumHealth — Agentic Healthcare Maps for India

Multi-agent healthcare intelligence system over 10,000 messy Indian facility records. Trust-scored, sentence-cited, geospatially aware. Built on the Databricks Data Intelligence Platform for Hack-Nation 5 — Serving A Nation challenge.

🔗 Live demo: https://qhealthtest.lovable.app

The Problem

In rural India, “static” healthcare data is unreliable. A facility might be listed as having an ICU but actually lack functioning equipment or 24/7 staffing. During the Golden Hour, this data gap costs lives. Existing tools surface raw lists; what’s missing is verified, trust-scored, agentic recommendations with reasoning a human can audit.

Our Approach — Three Pillars

Pillar 1 — Data & Retrieval Layer

• 10,000 messy Excel rows → cleaned Delta Lake table in Unity Catalog
• Mosaic AI Vector Search with HNSW index on 1,024-dim GTE-Large embeddings
• retrieve_for_agent(query, k, filters) returns Data Contract JSON in ~1.5s

Pillar 2 — Agentic Reasoning

• Auditor Agent (Llama-3.3-70B): extracts trust score, Truth Gaps, verified capabilities — with mandatory sentence-level citations
• Validator Agent (Llama-3.3-70B): independently reviews Auditor’s output, applies medical-standard rules, flags hallucinations
• Conservative ensemble: final trust = min(auditor, validator) + statistical confidence interval
• Parallel execution: per-facility audits run via ThreadPoolExecutor to stay under the SRS 10s budget

Pillar 3 — Crisis Command Center

• React frontend deployed on Lovable
• Leaflet + CartoDB Dark Matter tiles for the geospatial view
• Google Maps integration for turn-by-turn navigation
• Trust-graded pins (green/yellow/red) with severity-calibrated recommendations
• Slide-out “Truth Box” with sentence-cited reasoning trace
• Admin Mode: Desert Heatmap, Self-Correction Delta chart, Critical Supply Chain Needs

Architecture

Excel (10k rows)
    │
    ▼
[Pillar 1] Delta Lake → Mosaic AI Vector Search Index
    │
    │  retrieve_for_agent(query, k, filters)
    ▼
[Pillar 2] Lead Auditor (Llama 3.3-70B)
              │
              ▼
           Validator (Skeptic)
              │
              ▼
   {trust_score, truth_gap_notes, verified_capabilities,
    evidence_citations, confidence_interval, ...}
    │
    │  Databricks Model Serving REST endpoint
    ▼
[Pillar 3] React + Leaflet/CartoDB + Google Maps
   - Citizen Mode: trust-graded pins, Truth Box, Navigate
   - Admin Mode: Desert Heatmap, Self-Correction Delta,
                 Critical Supply Chain Needs

Data Contract

Every result returned by lead_auditor_agent() conforms to:

{
  "facility_id": "string",
  "facility_name": "string",
  "pin_code": 800020,
  "state": "Bihar",
  "city": "Patna",
  "coordinates": { "lat": 25.61, "long": 85.13 },
  "trust_score": 8,
  "verified_capabilities": ["NICU", "24/7 Emergency"],
  "truth_gap_notes": "[S2] claims Advanced Cardiac Care but [S4] lists no anesthesiologist.",
  "evidence_citations": ["[S1] 24/7 emergency care confirmed."],
  "is_medical_desert": false,
  "interval_label": "8.0 – 9.0 / 10",
  "confidence_note": "High — both agents aligned on evidence",
  "recommendation": "Best match — go here first",
  "agent1_score": 8,
  "validator_score": 9,
  "latency_ms": 5430
}

How to Run

Pillar 1 + 2 — Databricks Notebook

1. Open 02_retrieval_layer.ipynb in your Databricks workspace
2. Update the email path at the top: mlflow.set_experiment("/Users/<your-email>/serving-a-nation")
3. Upload the cleaned Parquet to a Unity Catalog Volume, then run the cells in order:
   • Cells 1–5: install deps + create Delta table + vector endpoint + index
   • Pillar 2 cells: define lead_auditor_agent() + validator_agent()
4. Deploy lead_auditor_agent as a Databricks Model Serving endpoint

Pillar 3 — Frontend

cd src/
npm install
npm run dev

Set in your .env:

VITE_AGENT_ENDPOINT=https://<workspace>.cloud.databricks.com/serving-endpoints/<endpoint>/invocations
VITE_DATABRICKS_TOKEN=<scoped PAT — model-serving + mlflow + vector-search scopes>
VITE_GOOGLE_MAPS_API_KEY=<your-google-maps-key>

Tech Stack

Layer	Technology
Data	Delta Lake, Unity Catalog, Change Data Feed
Retrieval	Mosaic AI Vector Search, HNSW, GTE-Large embeddings
Agents	Llama-3.3-70B via Databricks Foundation Model APIs
Observability	MLflow 3 Tracing
Backend serving	Databricks Model Serving
Frontend	React, Vite, Leaflet, CartoDB tiles, Google Maps API
Deployment	Lovable

Team

• Ami — Data pipeline & retrieval (Pillar 1)
• Urooj — Agentic reasoning engine (Pillar 2)
• Amaan — Crisis Command Center UI (Pillar 3)

Built for Hack-Nation 5 × MIT Club of Northern California × MIT Club of Germany — Serving A Nation challenge powered by Databricks.