Multi-Agent Healthcare Intelligence Platform for Ghana Virtue Foundation × Databricks × AI Tinkerers Hackathon — Bridging Medical Deserts
MedBridge AI analyses 797 medical facilities and NGOs across Ghana through 6 coordinated AI agents. Users ask natural-language questions and receive structured answers, interactive maps, and actionable insights — powered by a LangGraph orchestration pipeline with self-correcting feedback loops, optional quantum-optimised routing, and a cyberpunk-styled React frontend.
# Clone & install
git clone <repo-url> && cd MedBridgeAI
python -m venv .venv && .venv\Scripts\activate # source .venv/bin/activate on Linux/Mac
pip install -r requirements.txt
# Configure
cp .env.example .env # add GROQ_API_KEY, QDRANT_URL, QDRANT_API_KEY
# Run
uvicorn backend.api.main:app --host 0.0.0.0 --port 8000
cd frontend && npm install && npm run dev # → http://localhost:5173 User Question
│
▼
┌───────────┐ ┌────────┐ ┌────────────┐ ┌───────────┐ ┌──────────┐ ┌────────┐
│ Supervisor │───►│ Genie │ │ Vector │ │ Medical │ │Geospatial│ │Planning│
│ (Router) │ │Analyst │ │ Search │ │ Reasoning │ │Navigator │ │Strategy│
└───────────┘ └───┬────┘ └─────┬──────┘ └─────┬─────┘ └────┬─────┘ └───┬────┘
│ │ │ │ │
└────────────┴──────────────┴─────────────┴────────────┘
│
▼
┌──────────────┐ ┌───────────┐
│ Aggregator │────►│ Frontend │
│ + LLM Summary│ │ Results │
└──────┬───────┘ └───────────┘
│ retry if 0 results
└──────► (self-correction)
Flow modes: Sequential (e.g. Vector Search → Medical Reasoning) · Parallel (e.g. Genie + Geospatial) · Single agent
Classifies user intent and decides which agents to call.
- Top-2-mean embedding pooling — averages the 2 best similarities per intent (more robust than max-pool)
- Sigmoid confidence gating —
1/(1+exp(-20*(gap-0.05)))for sharp clear/ambiguous discrimination - Multi-intent expansion — detects secondary intents (similarity > 0.40) and merges agent sets
- LLM fallback — when embedding confidence < 0.45, Groq classifies with agent-name validation
| Intent | Example trigger | Routed to |
|---|---|---|
count |
"how many hospitals…" | Genie |
distance_query |
"near Kumasi", "within 30 km" | Geospatial |
validation |
"suspicious claims" | Vector Search → Medical Reasoning |
planning |
"deploy", "where should we build" | Planning |
coverage_gap |
"medical desert", "underserved" | Geospatial + Medical Reasoning |
comparison |
"compare Accra vs Northern" | Genie + Geospatial |
Structured Pandas queries over the full facility dataset.
- Counting, filtering, aggregation, region breakdowns
- Negation detection — "facilities without cardiology" correctly inverts filter masks
- IQR anomaly detection — adaptive
Q75 + 1.5×IQRthresholds for bed/doctor ratios - Returns records with lat/lng for map display
Semantic similarity search across Qdrant Cloud (384-dim, 3 named vectors per facility).
| Vector | Content | Query template |
|---|---|---|
full_document |
Complete facility profile | Raw query |
clinical_detail |
Specialties, procedures, equipment | "Procedures: {q} | Equipment: {q}" |
specialties_context |
Specialty names | "facility with specialties: {q}" |
- Reciprocal Rank Fusion (RRF) — merges results across all 3 vectors with normalised weights (sum = 3.0)
- City/region OR-filter for location-scoped queries
- Dual backend: Qdrant Cloud or Databricks Vector Search
Validates facility claims and detects data anomalies using medical domain knowledge.
| Mode | What it checks |
|---|---|
| Constraint Validation | Does a facility claiming neurosurgery have CT/MRI/ICU? |
| Anomaly Detection | Two-stage Isolation Forest + Mahalanobis outlier detection |
| Red Flag Detection | Language patterns suggesting exaggerated capabilities |
| Coverage Gap Analysis | Regions with zero or few providers for a specialty |
| Single Point of Failure | Specialties relying on only 1–2 facilities nationwide |
Distance calculations, coverage mapping, and medical desert detection.
- BallTree spatial index (Haversine) over 767 geocoded facilities — O(log n) radius/k-nearest queries
- Grid-based cold-spot detection — 0.25° grid across Ghana, flags cells > 50 km from any facility
- Medical desert detection — regions where citizens travel > 75 km to reach a specialty
- Mahalanobis distance — multivariate regional equity anomaly detection
- 3-stage geocoding — exact match → boundary check → fuzzy Levenshtein (handles "Kumase" → Kumasi)
Generates actionable deployment, routing, and resource allocation plans.
- Capability scoring — specialty match +35, ICU/theater +20, equipment +15 → routes by medical need
- 2-opt TSP — iteratively swaps edges to shorten specialist deployment tours (~15-20% reduction)
- Quantum QUBO routing (opt-in) — TSP as Quadratic Unconstrained Binary Optimisation via Qiskit; ≤4 cities use
NumPyMinimumEigensolver(exact), 5–10 use QUBO-aware brute-force; returns side-by-side comparison - Maximin placement — new facility GPS coordinates maximising minimum distance to existing facilities
| Scenario | Output |
|---|---|
| Emergency Routing | Primary + backup facility with distance/travel time |
| Specialist Deployment | Multi-stop optimised rotation route |
| Equipment Distribution | Priority list for underserved facilities |
| New Facility Placement | GPS coordinates for optimal new locations |
| Capacity Planning | Region-by-region status and expansion priorities |
React 19 · Vite 6 · Tailwind CSS v4 · DaisyUI 5 · Leaflet 1.9 — cyberpunk dark theme
| Tab | Content |
|---|---|
| ◈ Results | Colour-coded agent sections, facility tables with pagination, planning cards |
| 📋 Explain | Plain-language step-by-step explanation for NGO planners |
| ⟐ Trace | Agent timing, actions, confidence scores, LLM enhancement indicators |
| ◎ Map | Interactive Leaflet — colour-coded markers, dashed route lines, desert zones, proposed facility diamonds |
| ⚙ MLOps | Databricks pipeline status, MLflow tracking, model serving |
Key features:
- Markdown-rendered AI Summary — LLM output (bold, bullets, headings) rendered as formatted JSX via shared
renderMarkdownutility - Auto-tab-switch — geo/planning queries open the Map tab automatically
- Planning sidebar — 5 one-click scenario buttons (Emergency, Deployment, Equipment, Placement, Capacity)
- CSV export — one-click download of structured results
- Dynamic map legend — adapts to show facility types, desert severity, and proposed locations
| Component | Algorithm | Impact |
|---|---|---|
| Supervisor | Top-2-mean pooling + sigmoid confidence | ~15% fewer misroutes vs max-pool |
| Supervisor | Multi-intent expansion (> 0.40 threshold) | Complex queries activate all relevant agents |
| Geospatial | BallTree Haversine spatial index | O(log n) vs O(n) brute-force |
| Geospatial | Mahalanobis regional anomaly detection | Catches multivariate outliers z-scores miss |
| Planning | 2-opt TSP + QUBO quantum routing | Optimal specialist deployment tours |
| Planning | Maximin facility placement | Maximises geographic dispersion |
| Vector Search | RRF with normalised weights (sum = 3.0) | Balanced fusion across 3 vector types |
| Genie | Negation detection + IQR anomaly threshold | Handles "without" queries; adaptive cutoffs |
| Geocoding | 3-stage lookup (exact → boundary → fuzzy) | Handles typos safely |
| LLM | Token-aware truncation (binary-search slicing) | Prevents context overflow |
| Graph | Empty-result self-correction loop | Auto-retries without filters on 0 results |
MedBridgeAI/
├── backend/
│ ├── api/
│ │ ├── main.py # FastAPI app, CORS, lifespan
│ │ └── routes.py # /api endpoints
│ ├── agents/
│ │ ├── supervisor/agent.py # Embedding intent + multi-intent detection
│ │ ├── genie/agent.py # Pandas queries + negation + IQR anomalies
│ │ ├── vector_search/agent.py # RRF multi-vector search
│ │ ├── medical_reasoning/agent.py # Validation + Isolation Forest
│ │ ├── geospatial/agent.py # BallTree spatial + Mahalanobis equity
│ │ └── planning/agent.py # Capability scoring + 2-opt + QAOA + maximin
│ ├── core/
│ │ ├── config.py # Constants, specialty maps, API keys
│ │ ├── geocoding.py # 3-stage geocoding (exact/boundary/fuzzy)
│ │ ├── llm.py # Groq LLM + token truncation
│ │ ├── preprocessing.py # CSV → clean → dedup → geocode → documents
│ │ ├── vectorstore.py # Qdrant multi-vector + query templates
│ │ ├── quantum.py # QUBO TSP solver (Qiskit eigensolver)
│ │ └── databricks.py # Databricks Vector Search dual-backend
│ └── orchestration/
│ └── graph.py # LangGraph StateGraph + self-correction
├── frontend/src/
│ ├── App.jsx # Layout, tabs, query handling, data extraction
│ ├── utils/renderMarkdown.jsx # Shared markdown → JSX renderer
│ ├── api/client.js # API client
│ └── components/ # ResultsPanel, MapView, ExplainPanel, etc.
├── databricks/
│ └── medbridge_mlops_pipeline.py # MLflow + Delta tables notebook
└── data/
└── Virtue Foundation Ghana v0.3 - Sheet1.csv
| Method | Path | Description |
|---|---|---|
GET |
/api/health |
Health check |
POST |
/api/query |
Run a query through the full agent pipeline |
GET |
/api/facilities |
All facilities (for map markers) |
GET |
/api/stats |
Dataset statistics |
GET |
/api/specialties |
Available medical specialties |
GET/POST |
/api/planning/* |
Planning scenarios and execution |
GET/POST |
/api/mlops/* |
Databricks MLOps pipeline status and triggers |
| Category | Query |
|---|---|
| Counting | How many hospitals offer cardiology? |
| Negation | Facilities in Ashanti without orthopedic services |
| Proximity | Which facilities handle trauma near Kumasi? |
| Coverage | Where are the medical deserts in Ghana? |
| Validation | Find suspicious facility capability claims |
| Planning | Where should we deploy mobile eye care units? |
| Comparison | Compare Accra vs Northern Region healthcare |
| Resilience | Which specialties depend on a single facility? |
| Emergency | Plan an emergency route for a cardiac patient near Tamale |
| Quantum | Deploy a cardiologist near Accra (with use_quantum: true) |
GROQ_API_KEY=your_groq_api_key # Required — LLM synthesis
QDRANT_URL=your_qdrant_cluster_url # Required — vector search
QDRANT_API_KEY=your_qdrant_api_key
# Optional — Databricks MLOps
DATABRICKS_HOST=your_databricks_host
DATABRICKS_TOKEN=your_databricks_token
VECTOR_SEARCH_BACKEND=qdrant # or "databricks"| Layer | Technology |
|---|---|
| Frontend | React 19, Vite 6, Tailwind CSS v4, DaisyUI 5, Leaflet 1.9 |
| Backend | Python 3.11+, FastAPI, LangGraph |
| Vector DB | Qdrant Cloud (384-dim, 3 named vectors) / Databricks Vector Search |
| Embeddings | SentenceTransformer all-MiniLM-L6-v2 |
| LLM | Groq Cloud |
| ML | scikit-learn (Isolation Forest, BallTree), rapidfuzz |
| Quantum | Qiskit 2.3 + qiskit-optimization 0.7 |
| MLOps | Databricks (MLflow, Delta tables, Model Serving) |
| Data | Virtue Foundation Ghana CSV — 987 rows → 797 unique facilities |
MIT License