AI Voice Support Agent, built in week 6 of Koya Talent
A voice agent that answers support questions from approved knowledge and hands sensitive cases to a person.
Built with Next.js, Claude Agent SDK, MCP, Vapi, Supabase, Render
- A step in the flow
- Checked in code
- A person signs off
- Name and email
- Speech to text
- Search approved knowledge
- Score the evidenceChecked in code, not by the model
- Answer, clarify, escalate or decline
- MCP toolsChecked in code, not by the model
- Check before speakingChecked in code, not by the model
- Specialist takes overA person approves before anything moves on
- Ops console
- MCP tools
- MCP tools
- scenarios run on the live system
- scenarios run on the live system
- amounts or contact details spoken
- amounts or contact details spoken
Code decides whether an answer is allowed
The server scores what the knowledge search found. Tuned on 30 labelled questions, it went from letting 5 of 10 unanswerable ones through to none.
Haiku 4.5, chosen by measurement
Sonnet and Opus followed the same rules on six high-risk scenarios. Haiku replied fastest, at about a fifth of Opus's cost.
The live demo is offline for now. I used up my free Render instance hours, so the workspace is suspended until they reset at the start of November. I'm working on a fix.
Hosted on Render's free tier, so the first load can take up to a minute, and only one call runs at a time. The login is for the staff console at /ops.
Open the live app