CareKaki

Track: Open Track

Inspiration

Older adults often want independence while their families want reassurance. Repeated check-in calls can feel intrusive, yet a normal conversation can contain a small detail—a missed meal, medication uncertainty, loneliness, or dizziness—that a family needs to notice quickly. CareKaki is designed around a simple idea: care without constant calling.

What it does

CareKaki gives Grandpa a calm, voice-first check-in and gives family members a live view of only the moments that matter. It turns natural conversation into structured care signals for:

• medication taken or uncertain
• meals and drinks
• mood and loneliness
• health mentions
• urgent safety risk

The family dashboard moves from passive reassurance to a clear next action. In the urgent demo scenario, a possible double dose plus dizziness triggers a call-now state, preserves the exact important quote, and produces a concise family care note.

How we built it

The prototype uses Next.js and React. The primary voice path uses OpenAI Realtime over WebRTC for low-latency speech, transcription, and interruption handling. A server route creates short-lived Realtime sessions so the permanent API key is not exposed to the browser.

A deterministic signal layer converts conversation events into medication, meal, mood, loneliness, health, and urgent-risk cards. The guided demo mode provides a reliable, repeatable judge flow without claiming a live API session when no runtime key is configured. The current prototype stores only a minimal local session summary.

Challenges

The hardest design problem was balancing calm conversation with fast escalation. CareKaki should not alarm a family over ordinary small talk, but it must surface a credible medication or safety signal immediately. We also had to keep the senior interface extremely simple while giving family members enough context to act without overwhelming them with a full transcript.

Accomplishments

• A working two-phone prototype with synchronized senior and family views
• Actionable care signals instead of transcript overload
• Clear escalation from okay, to check-in-later, to call-now
• Exact quote preservation for urgent context
• A judge-safe deterministic demo plus a production OpenAI Realtime path
• Public deployment, source code, and a new 2:29 walkthrough video

What we learned

Voice UX for care requires more than a good model response. Turn-taking, repetition, interruption, accessibility, privacy, and failure states shape trust. We learned to separate conversational warmth from the decision layer: Grandpa gets a natural check-in, while family gets a compact operational summary.

What's next

Next steps include consent controls, encrypted multi-device sync, caregiver notification preferences, multilingual Singapore support, and carefully evaluated escalation thresholds. Any healthcare deployment would require clinical, privacy, and regulatory review.

Safety and privacy

CareKaki is a coordination layer, not a medical device. It does not diagnose conditions, recommend medication, dispatch responders, or replace emergency services. The demo uses fictional care data.

RLC submission and AI disclosure

This project began before RLC Hacks. The RLC organizers explicitly permitted submission of the pre-existing CareKaki project; it is disclosed here for transparency. This RLC entry includes a newly produced demo video and a verified public deployment.

OpenAI Realtime is a core product component for voice and transcription. AI-assisted development tools were also used during implementation and documentation. The guided video uses deterministic demo controls so the recorded behavior is reproducible.

Built With

• next.js
• openai-realtime
• react
• tailwind-css
• typescript
• vercel
• webrtc