convez.ai
MCP ServerIntegração & APIsAvançado
rdevarakond88

MedRecord — Prontuário Eletrônico Multi-Agent Offline-First

rdevarakond88/medrecord

Prontuário digital offline-first com orquestração de 7 agentes para clínicas indianas

Ver no GitHub
InicialPluginScore · de 0 a 100

Aplicativo móvel para clínicas semi-urbanas que gerencia prontuários de pacientes em ambiente offline-first, usando orquestração de 7 agentes Claude. Implementa busca de pacientes, histórico de visitas, documentação clínica estruturada e fluxo de consentimento conforme DPDP Act indiano, com sincronização automática quando conectividade retorna.

Instalação

$npx convez-ai rdevarakond88/medrecord

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
1
Forks
0
Acompanhando
0
Saúde
28%
Último commit
há 2 meses
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Clínicas rurais/semi-urbanas com internet intermitente
  • Documentação clínica rápida (< 60 seg por visita)
  • Gerenciamento de consentimento de pacientes multidobrador
  • Sincronização automática de registros médicos offline

Veredito

Projeto robusto que demonstra orquestração real de agentes Claude em produção médica. Excelente referência de arquitetura multi-agent, offline-first e conformidade regulatória (DPDP), mas requer expertise em React Native, SQLite e design de fluxos clínicos para implementar/estender.

README

A tradução para o português deste README ainda está na fila.

MedRecord

A multi-agent mobile app for offline-first patient health records — built with Claude Code, React Native, and a 7-agent orchestration pipeline.


What It Does

MedRecord is a mobile application designed for semi-urban Indian clinics. It enables doctors to record, store, and manage patient health records in an offline-first environment — meaning the app works fully even when internet connectivity is poor or intermittent.

Key capabilities:

  • Patient search by mobile number — fast lookup in high-volume clinic settings
  • Visit history — full clinical record with consent-gated access across doctors
  • New visit documentation — structured data entry designed for under 60 seconds per visit
  • Consent flow — patients grant or revoke access to their records (compliant with India's DPDP Act)
  • Auto-sync — visit data queued locally and synced automatically when connectivity resumes

Problems It Solves

ProblemSolution
Unreliable internet in rural/semi-urban clinicsOffline-first local storage with sync queue
Paper-based records — slow and error-proneDigital entry optimized for under 60 seconds per visit
PII exposure on shared clinic devicesConsent-gated data scoping + auth-layer enforcement
Cross-doctor data leakageIDOR-resistant architecture; strict patient ownership model
DPDP Act complianceConsent flow with explicit grant/revoke per doctor

Agent Architecture

This project was built using a 7-agent orchestration pipeline running in Claude Code. Each screen goes through every agent in sequence before being committed.

FLOW LEVEL
  PM Agent — validates flow, conducts pre-flight and post-flow reviews

SCREEN LEVEL (per screen)
  Builder Agent      — builds mockup, applies fixes, wires API/SQLite
  Persona Critic     — evaluates screen from 5 user perspectives (scored)
  Security Agent     — audits for PII exposure, IDOR, consent gaps
  QA Agent           — creates test plan, analyzes edge cases
  Backend Agent      — implements and deploys endpoints from api-contracts.md
  Integration Tester — validates full connected loop across screens

Every step runs in a fresh Claude Code session to prevent context overload. Session scope, agent handoff rules, and build mistakes are documented in AGENT_ORCHESTRATION.md and LESSONS-AND-RUNBOOK.md.


Screen Flows

  • D2 — Patient Search / Home: Entry point; doctors search by mobile number
  • D3 — Patient Detail / History: Full record view, consent status management
  • D6 — New Visit: Clinical notes capture, offline queue, auto-sync on reconnect
  • D9 / P4 — Consent Flow: Doctor requests access; patient grants or revokes

Tech Stack

LayerTechnology
FrameworkReact Native + Expo SDK 54
LanguageTypeScript
Local Storageexpo-sqlite (async, transactional)
StateZustand (global) + TanStack Query (server)
NavigationReact Navigation native-stack
NetworkingAxios/Fetch + ngrok (dev tunnel)
Offline Syncsync_queue with netinfo connectivity detection
Authexpo-crypto (randomUUID), session-scoped access

Key Lessons Learned

1. Human oversight is mandatory in agentic workflows Agents execute within scope precisely — but miss high-level product completeness (e.g., sign-up flows, recovery paths, retention policy design). A PM-layer review at flow start and end is non-negotiable.

2. Real-device testing is not optional Simulators and web previews will not surface native keyboard behavior, touch target sizing, or contrast errors. These only appear on physical devices.

3. Security must be enforced at the data layer Visual hiding (CSS opacity) is not security. Sensitive data must be stripped before it reaches components — not after.

4. Integration testing is a separate discipline Individual screen success does not guarantee system success. Cross-flow scenarios (doctor requests consent → patient grants → doctor view updates) must be tested as connected experiences, not isolated units.


Repository Structure

medrecord/
├── agents/                  # Agent prompt specs per role
├── backend/                 # API endpoints, Prisma schema, deploy config
├── src/                     # React Native screens and components
├── docs/                    # Project state, API contracts, UX specs
├── mockups/                 # Screen mockup files
├── AGENT_ORCHESTRATION.md   # Full agent pipeline and sequencing guide
├── CLAUDE.md                # Claude Code session instructions
├── LESSONS-AND-RUNBOOK.md   # Build mistakes, environment fixes, runbook
└── START-DEV.md             # Local dev setup and startup commands

Status

354 commits · 18 deployments · Active development

Built by Rajesh Devarakonda — AI Workflow Architect and TPM building agentic systems with Claude Code.

Ver o original no GitHub

Bom para

Relacionados