Intake

Patient-owned health intelligence workspace that turns fragmented medical records, clinician notes, and patient-reported context into a traceable knowledge graph.

Live demo

What it does

Intake helps patients organize longitudinal health information and ask grounded questions about it. The application:

1. Imports JSON EMR exports, clinician notes, and voice or chat intake.
2. Extracts structured facts and resolves duplicate clinical entities.
3. Builds an ontology-grounded graph with source provenance and confidence gates.
4. Surfaces timelines, risk alerts, trends, specialty reports, and citation-backed answers.

Technical highlights

• Knowledge graph: 12 node types and 13 validated relationship types with entity resolution, provenance edges, and corroboration-based confidence scoring.
• Agentic workflows: Grok-powered question-answering and trend agents use 11 tools for evidence retrieval, graph traversal, deterministic calculations, and structured final responses.
• Grounded answers: Inline citations resolve back to source events and graph nodes; low-confidence facts are rejected or flagged for review.
• Retrieval: BM25 lexical search with optional 384-dimensional embedding search and reciprocal-rank fusion through Supabase.
• Clinical rules: Ontology-backed medication, contraindication, interaction, laboratory, and trend rules complement model-generated analysis.
• Quality: 27 Vitest cases cover graph construction, entity merging, confidence scoring, provenance, and retrieval fusion.

Stack

• Application: Next.js 16, React 19, TypeScript, Tailwind CSS
• AI: xAI Grok chat, Responses, tool-calling, and realtime voice APIs
• Data: local-first browser storage with optional Supabase Postgres and Edge Functions
• Testing: Vitest

Architecture

EMR export / clinician note / voice intake
                    |
                    v
       parsing and AI fact extraction
                    |
                    v
 ontology grounding + entity resolution
                    |
                    v
 provenance-aware health knowledge graph
                    |
          +---------+---------+
          |                   |
          v                   v
 risk and trend rules   Grok tool-calling agents
          |                   |
          +---------+---------+
                    v
 timeline, graph, reports, and cited answers

Local setup

Requires Node.js 20 or newer.

npm install
npm run dev

Open http://localhost:3000.

AI features require an .env.local file:

XAI_API_KEY=your_xai_api_key
XAI_MODEL=grok-3-fast
XAI_TOOL_MODEL=grok-3-fast
XAI_VOICE_MODEL=grok-voice-latest

Supabase-backed persistence and semantic retrieval are optional:

NEXT_PUBLIC_SUPABASE_URL=your_project_url
NEXT_PUBLIC_SUPABASE_PUBLISHABLE_KEY=your_publishable_key

Without Supabase, account data persists locally in the browser.

Commands

npm run dev
npm run test
npm run lint
npm run build

EMR import format

Upload JSON containing any supported collection:

{
  "conditions": [{ "label": "...", "onset": "2024-01-01", "status": "active" }],
  "medications": [{ "label": "...", "dose": "...", "start": "2024-01-01" }],
  "labs": [{ "label": "...", "value": 5.2, "unit": "...", "date": "2024-01-01" }],
  "vitals": [{ "label": "Blood pressure", "value": "120/80", "date": "2024-01-01" }],
  "encounters": [{ "label": "...", "clinician": "...", "date": "2024-01-01" }],
  "careTasks": [{ "label": "...", "due": "2024-01-01" }]
}

Important limitations

Intake is an educational prototype, not a medical device. It does not diagnose conditions or replace professional medical advice. The portal connector included in the repository is a mock integration; EMR ingestion currently uses uploaded JSON exports.