AI system · 2026
Voice support agent
First-line voice support for a payments company. It answers from approved knowledge, checks records when it's safe, and hands the hard cases to a person.
- Role
- Design, build, deploy
- Stack
- Vapi, Claude Agent SDK, MCP, Supabase, pgvector, Next.js
- Status
- Archived
- Links
- Walkthrough (opens in a new tab)Code (opens in a new tab)App (opens in a new tab)Console (opens in a new tab)
- The recording didn't capture the agent's voice, though it speaks every reply.
Overview
A growing payments company couldn't keep up with support. Most questions were already answered in its docs. The rest touched someone's money and needed care.
Decisions
Every reply is checked before it's spoken.
Code compares each answer with what this turn's tools actually returned. Ticket numbers, times and dates are written by code, never by the model.
Two facts to check an account.
A caller needs two details that match one record. A miss and a partial match get the same reply, and contact details never leave the tool server.
A handoff books a real callback.
It books a slot in a real calendar and emails the team the whole call, so the customer never has to repeat themselves.
Architecture
- Person
- Tool
- Model
- Stored
- Check
- Result
Step 1, Human: A customer speaks or types
Caller
Step 2, Step: Speech to text
Vapi
Step 3, AI: Decide what to do
Sonnet 5.5, Agent SDK
Step 4, Data: Use the support tools
MCP server
Step 5, Logic: Check before speaking
Code
Then: step 3 (fails? repair once); step 6.
Step 6, Output: Spoken back
Vapi
Step 7, Data: On record for the team
Supabase, console
Results
22 / 24
test scenarios passed
59 / 60
benchmark runs on the chosen model
4.6 s
average turn on a live call
From the test runs and a live call.
Known limits
Every turn starts the agent fresh.
That start is still the biggest cost on a turn, and an always-on server would remove it.
Two matching facts isn't authentication.
A live version would send a one-time code to the email on file.