Inspiration
I currently live with my nana and great aunt, who both deeply struggle with technology. Even the basics of checking emails or adding events to a calendar are frequently insurmountable tasks. Because tech has become critical to their lives (getting updates from doctors, paying bills) they need a solution. And that is why I've built TechTalk.
What it does
TechTalk is a voice agent hosted by Vapi that works by simply calling a number. My agent connects to a range of MCP servers on Composio that gives TechTalk the agency to help the user with a wide variety of issues. It can check and send emails, search for important files in your drive, find calendar conflicts, and much more. People like my nana no longer need to struggle with the daunting task of attaching a file to an email. Just call TechTalk, say what you need, and it's done.
How we built it
The voice agent was created through Vapi using GPT 4o, a detailed system prompt, and access to a variety of tools through MCP servers. These servers were created through Composio, and currently use a suite of Google applications such as Google Drive, Google Calendar, and Gmail. My simple and intuitive web dashboard allows users to easily authenticate through their Google account and it's ready to go.
Challenges we ran into
When using STT --> LLM --> TTS workflows, latency is always a concern. Especially when you throw tool calls into the mix, there can be a large delay between question and answer. Since I'm working with some critical information like users personal gmails and files, I want to have guardrails and have users confirm actions before the agent takes them. It took some workshopping and detailed prompt engineering to find a sweet spot of safety and latency.
Accomplishments that we're proud of
Making it simple and varied. It's fairly easy to get connected and started, and the range of possibilities is quite large. Especially for something I built solo in just a couple hours, I think TechTalk is already off to a strong start. Not only do the tools work independently, but they can be combined to do tasks such as downloading a file a users drive and sending it in an email in one query.
What we learned
I learned a lot about Vapi and Composio through working with their tools. I've previously used Twilio and setup my own MCP servers, but these two providers made it so much easier and quicker to setup my project. Lots of overhead was reduced and I'll be sure to use them in the future.
What's next for TechTalk
Adding even more tools and preparing it for users. I have a simple dashboard that allows for oAuth, but there's no account authentication yet. The agent just uses the latest Google auth credentials. Adding tools such as Google maps and web search will make TechTalk even more adaptable and provide so much help to users.
Built With
- composio
- mcp
- next.js
- typescript
- vapi
Log in or sign up for Devpost to join the conversation.