CarePoint AI System - Build Instructions

System Overview

FastAPI endpoint receives text/image → LLM Council debates → Returns text response → Dashboards visualize

1. API Endpoint (FastAPI)

POST /consult
Input: { text: str?, image: base64?, patient_id: str, location: str }
Output: { response: str, urgency: str, confidence: float, council_votes: {} }

2. LangGraph Council (Core Intelligence)

Orchestrator: Routes based on input

• Text-only → Fast path (single GPT-4)
• Image → Visual path (Gemini multimodal)
• High-stakes → Full council (3 models debate)

Council Members (parallel execution):

• Gemini 2.0 Flash: Vision + text analysis
• GPT-4o: Fast reasoning
• Claude Sonnet 4: Medical expertise

Debate Phase: Agents critique each other Synthesis: Generate consensus response

3. LangChain Tools

• Medical knowledge API
• Drug database
• First aid instructions
• Emergency protocols

4. Arize Monitoring

from arize.otel import register
from openinference.instrumentation.langchain import LangChainInstrumentor

# Instrument all LLM calls
tracer_provider = register(space_id=..., api_key=...)
LangChainInstrumentor().instrument(tracer_provider=tracer_provider)

Logs: prompts, responses, latencies, council votes, confidence scores

5. Lovable Dashboards (Build in 1-2 hours)

Dashboard 1: Doctor Portal

Pull from Arize API:

// Fetch live consultations
GET https://api.arize.com/v1/spaces/{space_id}/traces

Display:
- Active consultations (real-time)
- Council voting breakdown (Gemini/GPT/Claude decisions)
- Model confidence scores
- Patient queue (waiting for doctor)
- [JOIN VIDEO CALL] buttons (LiveKit integration)

ARIZE_API_KEY=ak-b15b7f66-b95f-42f7-b5d8-a06b808b9ed7-b7gDKrO0sKTp27snEuQyQh5uh0Jsa2xe ARIZE_SPACE_KEY=U3BhY2U6MzU4NDA6M1BiWg==

Dashboard 2: Admin Analytics

Pull from Arize API:

// Fetch metrics
GET https://api.arize.com/v1/spaces/{space_id}/metrics

Display:
- Total consultations today
- Emergency detection rate
- Model accuracy comparison (bar chart)
- Average response time (line chart)
- Device status map (all first aid stations)
- Council agreement rate (pie chart: unanimous/majority/split)

Lovable Build Steps:

1. Sign in to lovable.dev
2. Prompt: "Create healthcare dashboard with real-time consultation cards, model performance charts, and device status map"
3. Add Arize API integration (paste API endpoint)
4. Customize: Add LiveKit room join buttons
5. Deploy: Get shareable URL for demo

6. Integration Flow

Text/Image Input
    ↓
FastAPI Endpoint
    ↓
LangGraph Routes
    ↓
Council Debates (if needed)
    ↓
Text Response Output
    ↓
[Parallel] → Arize Logs Everything
    ↓
Lovable Dashboards Pull & Display

Build Priority

Day 1:

1. FastAPI endpoint (text input only)
2. Single LLM response (GPT-4)
3. Arize basic instrumentation
4. Test end-to-end

Day 2: 5. Add LangGraph 3-model council 6. Add debate/synthesis phases 7. Add image input (Gemini) 8. Build Lovable dashboards 9. Connect dashboards to Arize API 10. Polish & test

Tech Stack

• FastAPI (API)
• LangGraph (orchestration)
• LangChain (tools)
• Arize (monitoring)
• Lovable (dashboards)
• OpenAI/Anthropic/Google (LLMs)