CareVoice Logo

About the Project — CareVoice

Inspiration

Medication adherence is one of the most persistent and preventable challenges in healthcare, especially for elderly populations. The problem isn’t a lack of concern or responsibility—it’s that most digital health tools assume users are comfortable with screens, typing, small text, and complex workflows.

We were inspired by a simple question: what if medication support didn’t require a screen at all?

Voice is the most natural interface for many elderly users. It removes cognitive and physical barriers and allows interaction in a way that feels familiar, calm, and human. CareVoice was born from the idea that AI should adapt to people—not the other way around.

What We Built

CareVoice is a voice-enabled AI assistant that helps elderly users confirm and track whether they have taken their medications—and whether they took them at the right time—using calm, natural conversation.

Key characteristics:

100% voice-based interaction — no typing, no reading
Confirms medication intake and timing (on time, early, late, or not yet taken)
Handles uncertainty with simple, reassuring clarification questions
Does not provide medical advice or alter prescriptions

The experience is intentionally minimal, supportive, and non-judgmental.

How We Built It

CareVoice is built using a lightweight, cloud-native architecture focused on reliability and accessibility.

Core Technologies

Google Cloud Run for serverless backend deployment
Google Vertex AI (Gemini) for intent understanding, reasoning, and clarification logic
ElevenLabs for high-quality speech-to-text and human-like text-to-speech
Lightweight session state (Firestore with TTL) for daily medication tracking and time-window checks

Architecture Overview

User voice input flows through a simple web UI into a Cloud Run backend, which orchestrates:

Speech-to-text via ElevenLabs
Reasoning and time-of-day validation via Gemini
Minimal session state updates for the current day
Text-to-speech via ElevenLabs for a calm audio response

The system is time-zone aware and uses predefined time windows (e.g., morning, evening) to determine whether medications were taken “on time,” without introducing complex scheduling logic.

What We Learned

Building CareVoice taught us several important lessons:

Voice UX is fundamentally different from screen UX.
Response length, pacing, and tone matter far more than feature richness.
Less logic in code, more clarity in prompts.
Carefully designed AI prompts reduced complexity and improved reliability.
Graceful failure is essential.
In voice applications, silence is worse than an error. Every failure mode needs a calm recovery path.
Intentional scope beats ambitious scope.
By explicitly excluding reminders, medical advice, and complex integrations, we delivered a stable and compelling MVP.

Challenges We Faced

Speech ambiguity:
Users often speak imprecisely (“I think I took it earlier”), requiring careful clarification logic that didn’t feel interrogative.
Latency management:
Combining speech processing and AI reasoning can introduce delays, so we optimized for short responses and lightweight state.
Time tracking without overengineering:
Determining whether a medication was taken “at the right time” required time-zone–aware logic and simple, human-friendly time windows—without building a full scheduling system.
Balancing safety and usefulness:
We had to ensure the system was helpful while clearly avoiding medical advice or prescriptive guidance.

Impact & Looking Forward

CareVoice demonstrates how voice-first AI can improve medication adherence while preserving dignity and reducing cognitive load. The same foundation can scale to caregiving, assisted living, and home healthcare environments.

Future extensions could include smart reminders, caregiver notifications, multi-language support, and deeper healthcare integrations—but this MVP proves that even a focused, simple approach can have meaningful real-world impact.

CareVoice is a step toward more humane, accessible AI in healthcare.