🏆 Winner — Best Design at H0: Hack the Zero Stack (Vercel v0 × AWS Databases).
One shared record for everyone caring for someone. The calm, role-aware home for a family's caregiving — across siblings, cities, and time zones.
Kintwadi (kin-TWAH-dee) is Kikongo for togetherness — doing it as one. Siblings, relatives, the parent, the hired aide, the clinician: one circle, one record.
🔗 Live demo → https://kintwadi.vercel.app
No sign-up needed: click "Run demo" on the homepage (or sign in as one of the seeded demo personas) to land in a living, RLS-scoped care circle on real Amazon Aurora PostgreSQL + Amazon Bedrock.
Built for the H0: Hack the Zero Stack hackathon (AWS Databases + Vercel/v0), where it won 🏆 Best Design.
Track: Monetizable B2C · Primary database: Amazon Aurora PostgreSQL (Serverless v2) · AI: Amazon Bedrock (Claude + Titan) · Frontend: Next.js on Vercel.
When an aging parent starts needing help, their care gets coordinated over chaotic group chats, sticky notes, and half-remembered phone calls — usually by one exhausted "default" child while siblings abroad feel guilty and out of the loop. Medications get missed or doubled. Nobody has the full picture.
Kintwadi is a single, shared, permission-aware care record for a family and their helpers. Everyone — the local daughter, the brother in another country, the hired aide, even the parent — sees the right slice of one source of truth: today's meds, the care timeline, appointments, tasks, vitals, and documents. An AI Daily Digest turns a day of small care logs into a warm human update for the relative who couldn't be there, and Ask Kintwadi answers natural-language questions grounded in the record — both running entirely on AWS.
Amazon Aurora PostgreSQL (Serverless v2) is the primary back end — and it's a thesis, not a default. Caregiving is an inherently relational, transactional, access-controlled domain (people, roles, permissions, medications, administration events, appointments, vitals, documents, audit). That demands joins, referential integrity, and atomic transactions — exactly what a relational engine guarantees and a key-value store does not.
The deliberate architectural choices Aurora makes possible:
- Row-Level Security (RLS) — RBAC enforced in the database. A hired aide physically cannot read a financial document; a member of one family cannot read another's record. Tenancy is a single indexed
circle_idcheck governed by Postgres policies (drizzle/0001_rls_policies.sqland the per-feature*_rls.sqlmigrations), not by trusting the UI. The app connects as a least-privilege role (carecircle_app) that is subject to RLS; only migrations/seed use the owner connection. - Atomic (ACID) medication safety. "Give a medication" writes the administration event, decrements supply, appends a timeline entry, and writes the audit row in one transaction (
src/lib/medications/actions.ts→recordDose). - Append-only audit log.
audit_loghas no UPDATE/DELETE policy — medical-grade, immutable accountability. pgvectorin the same database powers "Ask Kintwadi". Documents, timeline events, and audit entries are chunked, embedded with Amazon Bedrock Titan, and stored as vectors in Aurora'srag_chunktable (HNSW cosine). Retrieval runs inside the same RLS-scoped transaction, so the similarity search inherits the exact same permission and sensitivity boundary as the documents vault — there is no second vector-store ACL to keep in sync. Claude on Amazon Bedrock (Converse API) then writes a grounded, cited answer, and that read is audited. One database, relational and vector.- Serverless v2 scales with load (and toward ~zero when idle); the roadmap notes a path to Aurora DSQL for multi-region strong consistency, because the families we serve are global.
The rest of the stack: Next.js on Vercel (scheduler-agnostic, CRON_SECRET-gated cron routes drive the nightly digest and the daily care scans — refill sweep, missed-dose reconciliation, decline alerts; see "Scheduled jobs" below), Amazon S3 (documents & photos), Amazon SES/SNS (email + urgent escalation — the email layer in src/lib/email.ts is provider-pluggable: SES first, with SMTP/Resend fallbacks for local dev and SES-sandbox accounts), and keyless AWS access in production — Vercel's OIDC token is exchanged for short-lived STS credentials (AWS_ROLE_ARN), so no long-lived AWS keys are stored anywhere. Infrastructure is defined in Terraform under infra/.
A full diagram + component legend lives in docs/architecture.md (editable source: docs/Kintwadi-Architecture.drawio); the data model is documented in docs/data-model.md.
| Pillar | What it does |
|---|---|
| Shared care record | Recipient profile (conditions, allergies, blood type, directives), documents vault, emergency card |
| Medications | Structured regimens + schedules, give/skip/refuse logging, PRN caps, supply & refill alerts, a drug-interaction/allergy safety check (illustrative demo KB — see note) |
| Care timeline | One chronological feed of meds, vitals, notes, photos, incidents — with comments & reactions |
| Appointments | Shared calendar, assignment, pre-visit prep checklist, post-visit summary |
| Tasks & rota | Assignable/recurring tasks, weekly shift schedule, fair-share contribution view |
| Roles & permissions | Owner, family (full/limited), professional caregiver, read-only, care recipient, clinician — DB-enforced |
| Vitals & health | BP, glucose, weight, sleep, mood, HR with trend charts & per-circle alert safe ranges (DB-persisted; out-of-range readings post urgent alerts + SNS escalation) |
| Smart layer (AI) | Daily Digest (warm AI summary, translated into each member's own language — the aide reads it in Tagalog, the son in Dubai in English, cached one model call per language per day), Ask Kintwadi (permission-aware RAG), daily care scans (refill sweep, missed-dose reconciliation, 3-day decline alerts) |
| Notifications | Role/event-aware alerts; urgent escalation (a logged fall alerts coordinators instantly) |
| Emergency mode | One-tap shareable emergency card for EMS/ER |
| Admin (B2B) | Cross-tenant platform console (email-allowlisted, audited) for staff/agencies |
See AGENTS.md for the full, enforced invariants. In short:
- Every authenticated page lives under
src/app/(app)/whose layout runsrequireSession(). proxy.tsgates every route unless explicitly allowlisted (optimistic edge layer only).- All tenant data access goes through
withAuthedDb()— an RLS-scoped transaction. Aurora RLS is the final guarantee. - Server actions/routes re-check auth and authorize the specific action against the user's real role.
- Secrets stay server-side; passwords hashed; reset tokens hashed, single-use, short-lived.
- Every sensitive/state-changing action emits two trails: an operational
serverLog(...)and a durablerecordAuditEvent(...).
Frontend
Database & data layer
AI (all on AWS)
Auth & security
Cloud, infra & messaging
Quality & CI
In one line: Next.js 16 (App Router, Server Actions) · React 19 · TypeScript · Tailwind v4 + Radix UI · Drizzle ORM · postgres.js · Auth.js (NextAuth v5) · AWS SDK v3 (Bedrock Runtime, S3, SES, SNS) · pgvector · LangChain text splitters · Terraform · Vitest + RLS integration suite in GitHub Actions CI.
Requires Node 20+, an Aurora PostgreSQL (or any Postgres 15+ with the
vectorextension), and — for the AI features — Amazon Bedrock access.
npm install
cp .env.example .env # then fill in the values (see comments in that file)
npm run db:migrate # apply schema + RLS policies (uses MIGRATION_DATABASE_URL)
npm run db:setup-rls # create the least-privilege carecircle_app role + prove RLS, repoint DATABASE_URL
npm run db:seed # seed the living demo circle (6 weeks of meds, vitals, tasks, docs, an incident…)
npm run dev # http://localhost:3000Demo personas (all password Kintwadi123, printed by the seed): maria@kintwadi.demo (coordinator) ·
paolo@kintwadi.demo (remote family) · grace@kintwadi.demo (aide — reads the digest in Tagalog) ·
antonio@kintwadi.demo (care recipient) · rosa@kintwadi.demo (read-only) · chen@kintwadi.demo
(clinician — read-mostly clinical view) · admin@kintwadi.demo (platform admin → /admin).
Sign in with any of them from /sign-in. Set NEXT_PUBLIC_DEMO_MODE=1 to also show the homepage
"Run demo" button, which creates an anonymous guest account (RLS-scoped to only the demo circle, family
role, wiped on re-seed) and lands a reviewer on the live dashboard with zero sign-up — see src/lib/auth/demo.ts
for the fail-closed security model.
Key environment variables (full list + guidance in .env.example):
| Variable | Purpose |
|---|---|
DATABASE_URL |
Least-privilege app role — RLS is enforced through this |
MIGRATION_DATABASE_URL |
Admin/owner connection for migrations & seed (bypasses RLS) |
AUTH_SECRET |
Auth.js session secret |
AWS_REGION, AWS credentials |
Bedrock / S3 / SES / SNS |
BEDROCK_MODEL_ID |
Claude inference-profile id (account/region-specific) |
CRON_SECRET |
Shared secret for the digest cron (Vercel injects it as Authorization: Bearer …) |
PLATFORM_ADMIN_EMAILS |
Allowlist for the cross-tenant /admin console |
- Import the repo into Vercel and set the env vars above in Project → Settings → Environment Variables (including
CRON_SECRET). - Aurora must be reachable from Vercel functions (publicly accessible instance or a tunnel/VPC connector).
- Scheduled jobs. Two
CRON_SECRET-gated, scheduler-agnostic routes: the Daily Digest (/api/cron/digest, pinged hourly — generates + emails each circle's digest at its localdigestHour, idempotent per day) and the care scans (/api/cron/scans, pinged every 6h — refill sweep, missed-dose reconciliation, decline alerts; every sweep idempotent). The schedules fire from GitHub Actions (.github/workflows/cron.yml) — a deliberate choice: Vercel's Hobby plan caps cron jobs at once per day, which can't drive an hourly, per-timezone digest. The routes accept any pinger holding the bearer secret, so on Vercel Pro the same endpoints move intovercel.jsoncrons with zero code change. - Recommended — keyless AWS via OIDC: trust Vercel's OIDC issuer in AWS IAM, set
AWS_ROLE_ARN, and remove the static AWS keys (seeSETUP.md§ "Vercel OIDC → AWS"). All AWS clients (Bedrock, S3, SES, SNS) resolve credentials throughsrc/lib/aws/credentials.ts.
npm test # unit tests (no AWS, no DB) — always run
TEST_DATABASE_URL=... npm test # additionally runs the RLS integration suite- Unit suite (
src/**/__tests__/*.test.ts): proves the application-layer authorization matrices (who can manage meds, record doses, see restricted documents, report/resolve incidents, post private notes, reach the cross-tenant admin console, …) and that the AWS Bedrock client is mocked — no real AWS calls in tests. - RLS integration suite (
src/db/__tests__/rls.integration.test.ts): runs only whenTEST_DATABASE_URLpoints at a throwaway Postgres. It applies the real migrations, connects as the least-privilegecarecircle_approle, and asserts the database itself enforces cross-tenant isolation and the sensitivity tiers forrag_chunkretrieval (a read-only member can never retrieve arestrictedchunk; circle A never sees circle B). AWS is never touched — vectors are inserted directly. - CI (
.github/workflows/ci.yml): every push runs lint, type-check, the unit suite, and the RLS integration suite against a realpgvector/pgvectorPostgres service — so the badge at the top of this README is a continuously re-verified proof of the security claim, not a one-time assertion.
This is a hackathon demo. Clinical content (e.g. the drug-interaction safety check in src/components/medications/safety.ts) uses a small illustrative knowledge base, clearly labeled in code — a production system would call a licensed clinical-interactions service (shown as an external dependency in the architecture diagram). All seeded people and records are fictional.