KinKeep

A family health companion that notices meaningful changes, checks in with your parent, and tells you when you're needed.

Agents own the execution. Families retain judgment and control.

Important

KinKeep is not a medical diagnostic tool. All health signals and care scenarios in this demo use synthetic fixtures. Care branching, escalation stages and approval gating are auditable code — never model output. No clinic or emergency service is contacted. Family messages are never sent automatically: KinKeep can open a pre-filled WhatsApp handoff, but the user must choose the recipient and press Send.

---

🎯 The idea

Every family has a kinkeeper — the one who remembers the pills, makes the daily call, coordinates everything. Apple Health can already share Mum's heart rate with her. But numbers are not care: they don't ask how Mum's knee feels, don't notice that she slept two hours less and walked only 320 steps and skipped breakfast on the same morning, and don't tell a busy daughter whether today is a day that needs her.

KinKeep closes that loop:

Personal change detection        Conversational check-in       One decision for the family
multi-signal, vs HER own    →    in her language, she just  →  first round of care completed;
baseline — never one number      answers, never initiates      humans keep the judgment

The elder gets a companion that talks like a person. The family gets silence when all is well — and a completed first round of checks plus one clear decision when it is not.

🧭 OpenAI Build Week scope and project history

KinKeep predates OpenAI Build Week: an earlier prototype was created during BUIDL_QUESTS 2026, which has now concluded. The current submission and development focus is OpenAI Build Week 2026. Because Build Week allows builders to extend an existing project, this repository preserves a clear before-and-after boundary instead of presenting the whole product as new.

What existed before Build Week

• A bilingual parent conversation and family dashboard
• A simulated Apple Watch morning summary and personal-baseline trends
• Basic symptom check-ins, consent prompts, approval/audit UI, and a synthetic wandering/escalation demo
• Optional server-side OpenAI companion replies, meal-photo analysis, and voice transcription
• The vinext/Cloudflare deployment scaffold, production build, lint, and SSR tests

What I added during Build Week

• A complete Care Episode Loop: visible health-data review → proactive bilingual inquiry → evidence and uncertainty → three care options → family approval → named owners and deadlines → follow-up → recorded outcome
• Shared cross-surface episode state so the parent's consent and check-in become the family's notification, decision, task plan, and resolved follow-up
• Consent-led family contact that asks the parent who to contact, creates one clear alert for that person, and lets the family acknowledge and call back
• Safer, more accessible check-ins with time-aware meal prompts, knee/stomach/dizziness danger checks, Singapore emergency and NurseFirst guidance, and optional read-aloud advice
• Submission-quality consistency and verification: unified synthetic fixtures, responsive card fixes, expanded rendered-output assertions, and end-to-end demo rehearsal

I used Codex to implement, refactor, debug, test, and prepare the working demo. I remained responsible for product scope, safety rules, review, integration, and every consequential care boundary.

Verifiable new-vs-existing boundary

Evidence	Reference
Last commit before OpenAI Build Week opened	10970be
Core Build Week Care Episode Loop	d1d1c74
Consent-led contact refinement and expanded tests	6cdac5a
Current final Build Week source on main	5ca6f83 — merged through PR #2
Full Build Week diff	10970be...main
Archived pre–Build Week prototype snapshot	judged-2026-07-12
Archived prototype build timeline	docs/BUILD_LOG.md

The current main includes the final Build Week source through PR #2; the history above separates it from the pre–Build Week baseline.

▶️ Recommended 3-minute demo route

Keep the story on one completed care episode:

1. On /, sync the simulated health summary.
2. Choose My knee hurts → A few days → None of those.
3. Show that KinKeep asks permission before contacting family, then choose Yes, please.
4. Switch to /family and open Today's follow-up.
5. Show the four evidence items, the visible uncertainty, and the three care options.
6. Approve Ask family to call and show the generated owner, deadlines, follow-up, and escalation rule.

Voice, meal-photo analysis, WhatsApp handoff, and wandering monitoring are supporting capabilities. Mention them briefly; do not spend the core three minutes demonstrating them.

✨ One product, two surfaces

Try it live: kinkeep-family-health.leini9591.chatgpt.site — parent at /, family at /family. Bilingual (EN / 简体中文) everywhere, switchable at runtime.

📱 / — the parent's side: it opens talking

💬 Conversation-first	No dashboard, no charts. Quick-reply chips instead of typing — she answers, she never has to initiate
⌚ Watch summary in the chat	Tap-to-sync Apple Watch morning card: sleep −1h · steps low · vs her own usual (simulated stream today)
🩺 Structured check-ins	Time-aware bilingual prompts for knee pain, stomach pain, sleep, appetite, and dizziness, with danger-sign checks before routine follow-up
🔊 Accessible safety guidance	Read-aloud Singapore guidance plus explicit choices for whether and how a family member should be contacted
🍜 Real meal-photo analysis	Server-side OpenAI vision: calorie range, macros, fibre/sodium/veg, portion, visible uncertainty, one practical suggestion
🎙️ Real voice input	In-browser recording → server-side transcription (zh/en), browser speech recognition where supported
🆘 "I need help"	Offers immediate guidance, then lets the parent choose the nearby responder David, Elena, or no family contact

👨‍👩‍👧 /family — the family's side: an answer, not a feed

✅ "Is Mum okay?" answered	One sentence with plain-language reasoning — evidence bullets, not raw data
👥 Relationship-aware profiles	Mum · Dad · self · brother — each with a scoped permission level, not one feed for everyone
📈 Trends & evidence	7-day activity / sleep / resting HR / routine-match, each drawn against her personal range band, with "why today was flagged"
🤝 Approval gate	High-impact actions (e.g. suggesting telehealth) pause with evidence + the matched policy reason; approve/decline both produce honest outcomes
🗓️ Care schedule & roles	Family + companion tasks on one timeline; weekly assignments across Elena, brother, and the companion
📜 Audit & access	Who viewed, who decided, what the system did — including "Mum approved sharing this reply": consent runs both directions
🚨 Escalation awareness	The selected channel and acknowledgement state sync across both surfaces, with synthetic location/HR/battery context for the wandering fixture

🛡️ The sovereignty model

Most agent demos show how far autonomy can go. KinKeep focuses on where autonomy must stop in family care.

Agents do autonomously	Pauses for a human
Read wearable summary; compare vs personal baseline	Choosing whether and which family member to contact — the elder decides
Initiate routine check-ins and prepare follow-ups	Contacting a family member — the parent chooses who receives the alert
Structured symptom triage and record-keeping	Sharing the elder's reply with the family circle — logged authorization
Meal understanding + one practical suggestion	Third-party data sharing, care-plan changes
Family summaries, options, tasks, and audit entries	Care-plan approval and emergency response — human acknowledgement, never auto-dial

🚀 Quick start

Requires Node.js ≥ 22.13.

npm install
npm run dev

Parent and family experiences run from the same server. Optional — for live meal vision and voice transcription, create an ignored .env.local:

OPENAI_API_KEY=your-key
# Optional override; defaults to gpt-5.6-terra
OPENAI_MODEL=gpt-5.6-terra

# Optional: international-format digits only, used to open the selected family chat
KINKEEP_ELENA_WHATSAPP=6591234567
KINKEEP_DAVID_WHATSAPP=6597654321

If a family WhatsApp number is not configured, KinKeep still opens WhatsApp with the message pre-filled, but the elder must select the recipient. The browser cannot press WhatsApp's final Send button on the user's behalf.

No OpenAI key? Everything else is deterministic and works offline — the demo never depends on a model to stay safe.

✅ Validation

npm run lint
npm test     # production build + server-rendered assertions for both surfaces

🧠 Built with + technical implementation

• Application: Next.js 16 App Router · React 19 · TypeScript · Tailwind 4
• Runtime and deployment: vinext/Vite on Cloudflare Workers, deployed through OpenAI Sites
• GPT-5.6: OpenAI Responses API with gpt-5.6-terra for short, grounded companion replies and schema-constrained meal-photo understanding; requests are server-side with store: false
• Voice: OpenAI Audio Transcriptions with gpt-4o-mini-transcribe; optional browser speech synthesis reads safety guidance aloud
• Deterministic safety layer: typed state machines control symptom branching, consent, escalation channels, approvals, owners, deadlines, and follow-up status — model output cannot bypass them
• Cross-surface demo state: typed CareEpisode and CareEscalation records persist in local storage and dispatch browser events so parent and family views stay aligned without pretending the demo has production clinical infrastructure
• Structured uncertainty: meal analysis uses strict JSON Schema output with confidence and assumptions instead of false precision
• Codex development workflow: milestone planning, code generation and refactoring, regression tests, debugging, responsive verification, and deployment preparation, with human review and final product judgment

📚 Documentation and evidence

• OpenAI Build Week scope and evidence: this README and the 10970be...main diff
• Three-minute feature pitch and click script: docs/OPENAI_BUILD_WEEK_3MIN_PITCH.md
• Archived pre–Build Week prototype timeline: docs/BUILD_LOG.md

🗺️ Roadmap

Real mobile push & notification ladder → real wearable ingestion (HealthKit / rings) → dialect voice (Cantonese · Hokkien · Teochew · Tamil) validated with real elders → Singapore hawker-food nutrition depth → Care Provider Console: one care professional, supported by agents, coordinating wellness for dozens of seniors.

---

OpenAI Build Week 2026 · built with Codex + GPT-5.6

Synthetic demo data · Non-diagnostic · Consent-first · Humans decide high-impact actions

KinKeep — notices the change, checks in first, tells you when you're needed.