Bedside

AI-powered inpatient assistant delivered via WhatsApp. Patients text like they text a nurse. Hospitals stay in the loop.

Harvard HSIL Hackathon 2026.

Bedside is a multi-tenant SaaS platform that lets hospitalized patients get trustworthy answers about their own care — schedule, medications, what's next, how to reach a nurse — by sending a WhatsApp message. The bot speaks the patient's language (pt-BR, es, en), grounds every answer in the hospital's own care plan, and escalates to clinical staff when something isn't safe to answer alone.

Hospital staff see everything from a dashboard: live patient list, pending escalations, audit trail, and direct patient messaging.

Why

Inpatients forget or misunderstand a large share of what they're told at the bedside. They don't want to press the call button for "what pill is this" or "when is my X-ray." Family members are anxious and under-informed. Nurses are interrupted constantly.

Bedside gives patients a safe, grounded way to ask — and gives clinicians a full record of what was asked and answered, without adding a single new app on the patient's side. WhatsApp is already on every phone.

What it does

For patients (WhatsApp):

• "What's on my schedule today?" → interactive list of appointments with time and location
• "What are my medications?" → interactive list with dose, route, and reason
• "What's next?" → the most urgent of upcoming appointment vs. medication dose
• "I need a nurse" / "dor" / "emergency" → routes to the escalations table, care team notified
• "family on" / "family off" → toggles a view-only share link for the family
• Free-text conversation grounded in the patient's care plan via the AI layer
• Auto-detected language: Portuguese (pt-BR), Spanish (es), English (en)

For hospital staff (React dashboard):

• Live patient list with ward, bed, condition, attending physician
• Pending escalations with one-click resolve
• Per-patient view: schedule, medications, consent flags
• Direct WhatsApp message to the patient, logged to nurse_messages
• Add patients, appointments, medications
• Audit log: every inbound intent, handler used, response summary, detected language — never the raw message content (LGPD-compliant)

Architecture

┌───────────────┐   webhook    ┌──────────────────┐
│  Evolution    │  ───────────>│  bedside-backend │
│  API (WA)     │<─ responses ─│  Fastify + Bun   │
└───────────────┘              └────────┬─────────┘
         ▲                              │
         │                              ▼
  patient phone              ┌──────────────────┐
                             │    Supabase      │
                             │  Postgres + RT   │
                             └────────┬─────────┘
                                      │ realtime
                             ┌────────▼─────────┐
                             │ bedside-frontend │
                             │  React + Vite    │
                             └──────────────────┘

Piece	Stack
Backend	Fastify · TypeScript strict · Bun · Supabase-js · Evolution API · model-agnostic AI (MistralAI wired in)
Frontend	React 18 · Vite · Tailwind · Supabase Realtime · Recharts
Database	Supabase (Postgres) — 8 tables, multi-tenant on hospital_id
WhatsApp	Evolution API v2.3.7 (Baileys) — docker-compose.yml included

Repo layout

bedside/
├── bedside-backend/          # Fastify webhook + handlers + admin API
│   ├── src/
│   │   ├── routes/           # webhook, admin, family
│   │   ├── handlers/         # schedule, medication, next_action, escalation, family, free_text
│   │   ├── ai/               # model-agnostic client, promptBuilder, languageDetector
│   │   ├── whatsapp/         # Evolution API sender, interactive message builder
│   │   ├── session/          # in-memory conversation history per patient
│   │   └── utils/            # patient lookup, audit logger, injection detector, phone normalizer
│   ├── seed/                 # schema.sql + seed.ts (Hospital Isaac Newton + 3 demo patients)
│   ├── docker-compose.yml    # Evolution API + Postgres + Redis
│   └── .env.example
├── bedside-frontend/         # React dashboard
├── docs/
│   ├── AGENT_1_BACKEND_v2.md     # Backend spec (source of truth)
│   ├── AGENT_2_FRONTEND_v2.md    # Frontend spec
│   ├── SCHEMA_v2.md              # Database schema (source of truth)
│   ├── GOOGLE_STITCH_PROMPT_v2.md
│   └── evolution-api-setup-guide.md
└── README.md

Quickstart — backend

Prerequisites: Bun, Docker, a Supabase project, and an AI provider key (MistralAI by default, any OpenAI-compatible provider via env).

cd bedside-backend
bun install
cp .env.example .env           # fill in Supabase + AI + Evolution API creds

Apply the schema (run seed/schema.sql in the Supabase SQL editor), then seed demo data:

bun run seed                   # Hospital Isaac Newton + 3 patients + appointments + meds

Start WhatsApp and the server:

docker compose up -d           # Evolution API on http://localhost:8080
bun run dev                    # Fastify on http://localhost:3000

Point Evolution API's webhook at http://<your-host>:3000/webhook/evolution and scan the QR code. Text the connected number from one of the seeded patient phones and you're in.

Quickstart — frontend

cd bedside-frontend
bun install
cp .env.example .env           # Supabase url + anon key
bun run dev

Deploy on Vercel

Use two Vercel projects from this monorepo:

• bedside-backend
• bedside-frontend

The deployment runbook is in docs/DEPLOY_VERCEL.md.

Safety, compliance, and design choices

• Multi-tenant isolation — every operational query filters by hospital_id. No cross-tenant leakage.
• LGPD-compliant audit logs — audit_logs store intent-based summaries only. Raw message content never leaves conversation_logs.
• Clinical safety on escalation — if the escalation insert fails, the patient is told to press the nurse call button or call out. The bot never gives false reassurance.
• Prompt injection detection — 8-pattern filter before the AI layer; hits are logged and answered with a safe fallback, the AI is never called.
• Word-boundary intent matching — "painter" doesn't trigger a pain escalation, "helpful" doesn't trigger a help escalation.
• Model-agnostic AI layer — generateResponse(messages: ChatMessage[]) behind an env-configured provider. No model name appears in patient-facing text.
• Webhook dedup — same message ID within 30s is skipped, entries cleaned after 60s.
• Session memory cap — 20 messages per patient, drop-oldest-2 on overflow.

Spec docs

The docs/ directory is the source of truth. Both the backend and frontend were built by reading these end-to-end.

• docs/AGENT_1_BACKEND_v2.md — backend spec
• docs/AGENT_2_FRONTEND_v2.md — frontend spec
• docs/SCHEMA_v2.md — database schema
• docs/evolution-api-setup-guide.md — Evolution API setup guide

Status

• Backend — 24 files, 15 build steps, typecheck clean, architect-reviewed
• Frontend — dashboard, patient detail, and escalations surfaces shipped
• Database — 8-table schema with realtime on audit_logs, escalations, conversation_logs, nurse_messages, patients