Inspiration
The physical world still relies heavily on phone calls for everyday tasks: booking appointments, checking deliveries, cancelling services, requesting quotes, and relaying important information.
For Deaf, hard-of-hearing, and speech-impaired people, these interactions can become a major barrier. Even when digital services exist, many businesses still require customers to call and navigate phone systems.
We were inspired by a simple question:
What if someone could access a phone-based service without personally having to speak or listen on the phone?
YourVoice was created as an accessibility voice proxy: the user communicates through text, while CALL-E handles the actual telephone conversation. The goal isn't to replace the user or hide the AI — it is to give the user more autonomy, transparency, and control.
What it does
YourVoice turns a user's written request into a controlled phone workflow using CALL-E.
A user can choose a guided task such as:
- Booking or rescheduling an appointment
- Checking an order or delivery
- Cancelling a service
- Requesting information from a business
- Checking in with an elderly family member
- Relaying an important message
- Comparing multiple vendors
For custom requests, YourVoice uses a two-phase planning flow. The system asks clarifying questions, creates a proposed calling plan, and lets the user review and edit it before anything is dialed.
After approval, CALL-E makes the phone call and returns structured information. YourVoice turns that information into a clear written outcome so the user doesn't have to rely on hearing or speaking during the call.
For multi-vendor tasks, YourVoice can contact multiple providers with the same questions and use AI to compare their responses.
Most importantly, the user remains in control: the system shows who will be contacted, what will be asked, and requires approval before placing a call.
How we built it
YourVoice is built as a full-stack Next.js application using TypeScript and React.
The core architecture combines:
- CALL-E Voice AI SDK — performs the actual outbound telephone calls and handles voice conversations.
- Model Context Protocol (MCP) — powers the interactive planning flow for ambiguous or open-ended requests.
- Google Gemini — analyzes and compares responses for multi-vendor workflows.
- Supabase — handles authentication, sessions, and persistent application data.
- Upstash Redis — provides rate limiting, daily call quotas, and idempotent webhook processing.
- Zod — provides strict validation for call inputs and structured task schemas.
The basic flow is:
User writes a request → YourVoice creates a plan → User reviews and approves → CALL-E places the call → CALL-E returns structured results → YourVoice presents the evidence and outcome.
We also designed the application around accessibility rather than treating accessibility as an afterthought. The interface uses high contrast, simple typography, minimal visual noise, written call outcomes, and explicit confirmation gates.
Challenges we ran into
One of the biggest challenges was designing a phone experience where the person using the application doesn't need to personally participate in the phone conversation.
A normal voice assistant can simply listen to the user and respond. That doesn't solve the accessibility problem we were targeting. We needed to create a reliable bridge between a text-based user interface and a real-time telephone conversation.
Another challenge was deciding when an AI should act and when it should stop and ask the user.
For example, an ambiguous request should not immediately result in a real phone call. We therefore built a planning stage that can ask clarifying questions and produce an editable task before execution.
We also had to think about the risks of real-world phone actions. A mistaken call can have consequences, so we implemented explicit approval, call quotas, validation, safe demo paths, and idempotent webhook handling.
Finally, turning unstructured conversations into useful, trustworthy information was challenging. We wanted the result to be more than a transcript, so the application extracts the important evidence and presents it as a structured outcome.
Accomplishments that we're proud of
We're proud that YourVoice is more than a concept or a UI mockup.
CALL-E is actually part of the runtime workflow. The application can formulate a task, obtain user approval, dispatch the call through the CALL-E SDK, receive the completion webhook, and present the resulting information back to the user.
We're particularly proud of the human-in-the-loop design.
The user can see:
What is going to happen → Who will be contacted → What will be asked → Approval → What happened → What was learned → What to do next
We are also proud of the multi-vendor workflow. Instead of making a single phone call, YourVoice can coordinate multiple calls with consistent questions and synthesize the responses into an explainable comparison.
Most importantly, we built the project around a specific accessibility problem rather than building a generic "AI that makes phone calls."
What we learned
We learned that accessibility isn't just about making an interface easier to use.
Sometimes the inaccessible part is the underlying system itself.
A beautiful web interface doesn't solve the problem if the final step still requires a person to navigate an auditory phone system.
We also learned that giving an AI the ability to take real-world actions changes how the product needs to be designed. Transparency, confirmation, limits, validation, and recoverability become just as important as the AI itself.
Finally, we learned that the best AI workflow is not always the one with the most automation. For sensitive or consequential actions, giving the user a clear opportunity to understand and approve what will happen can make the system substantially more trustworthy.
What's next for YourVoice - Accessibility Voice Proxy
Our next goal is to make YourVoice useful beyond the hackathon.
We want to expand the number of real-world phone workflows and make the planning system more adaptive while keeping the same human-control principles.
Future directions include:
- More accessibility-focused workflows for healthcare, government services, utilities, and customer support.
- Better support for regional languages and localized phone interactions.
- Persistent user preferences that reduce repetitive setup while preserving confirmation for consequential actions.
- More sophisticated call-result verification and evidence extraction.
- Better handling of IVR systems and multi-step phone menus.
- A stronger mobile experience for users who rely primarily on their phones.
- Accessibility testing with real users who are Deaf, hard of hearing, speech-impaired, or have other communication barriers.
The long-term vision is simple:
If a service is available through a phone call, the inability to speak or hear should not automatically mean the inability to access that service.
YourVoice is our attempt to build that bridge.
Log in or sign up for Devpost to join the conversation.