SupConnect: Supplier Voice Agent

An inbound AI voice agent that answers buyer calls for secondhand clothing suppliers, qualifies the deal inside supplier rules, and hands over a ready-to-action lead.

It is 3am in Karachi. A buyer in London wants 200 pieces of 90s denim. Normally that call rings out and the sale walks. This agent picks up, answers only from the supplier's own catalogue, qualifies the request, and leaves a structured lead in the dashboard by morning.

The agent qualifies. The supplier closes.

Screenshots

Live call	Lead the supplier wakes up to

What it does

Answers buyer calls 24/7 in the browser, live voice or text, on the same pipeline.
Answers only from supplier-approved knowledge. It never invents a price or a promise.
Fills the qualification fields live as the buyer talks, then confirms them back.
Escalates to a human when a rule says so, for example a discount ask or a complaint.
Writes a lead card with a grounded summary, transcript, and a recommended next action.

How it works

One rule runs everything: the model talks, deterministic code decides. The LLM never sets a lead status, never invents a number, and never chooses to escalate. When the call ends, the summary agent narrates the record the core already built. It rewords, it does not decide.

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#F6C445','primaryTextColor':'#0A0A0A','primaryBorderColor':'#C68A16','lineColor':'#F6C445','clusterBkg':'#141414','clusterBorder':'#333333'}}}%%
flowchart LR
  buyer([Buyer]) -->|voice or text| session[Agent session]
  session <-->|tool calls, facts, rules| core[Qualification core]
  core -->|live field updates| web[Supplier dashboard]
  core -->|lead record| summary[Summary agent]
  session -->|transcript| summary
  summary -->|lead card| web
  web --> supplier([Supplier])

For live voice, the qualification core runs in a session server (apps/server) and streams field updates to the browser over a WebSocket, so the chips fill in real time during the call.

The deciding is done by four parts:

Part	What it does	Status
Qualification state machine	Tracks filled fields, picks the next question, moves the lead between states	Built and tested
Escalation rules engine	Supplier-editable rules, checked on every turn	Built and tested
Knowledge lookup	Returns catalogue facts, or `not_found` so the agent says it does not know	Built and tested
Numeric provenance guardrail	Rejects any number in the summary that did not come from a knowledge lookup, then falls back to a deterministic template	Built and tested

Post-call summary

The summary agent turns the finalised lead into a short brief plus a few grounded key points, each with its own label. The status and next action come from the deterministic core, so the card renders even with no LLM key. On any provider problem, no key, timeout, bad JSON, or an invented number, it degrades to a deterministic template and the card still renders.

The eval harness (@fleek/evals) drives the pipeline through a set of buyer personas and asserts the agent stayed inside the rules: no ungrounded numbers, correct escalations, correct terminal status. It can run against a standalone scripted pipeline or the live server, and ships deliberately broken streams to prove the assertions catch a lying agent.

Lead lifecycle

%%{init: {'theme':'base','themeVariables':{'primaryColor':'#F6C445','primaryTextColor':'#0A0A0A','primaryBorderColor':'#C68A16','lineColor':'#F6C445'}}}%%
stateDiagram-v2
  [*] --> in_progress
  in_progress --> qualified_follow_up: contact and a real requirement, confirmed
  in_progress --> human_handoff_requested: an escalation rule fires
  in_progress --> unresolved: call ends before qualifying
  qualified_follow_up --> completed: supplier acts
  human_handoff_requested --> completed: supplier acts

Run it

The project is a pnpm workspace. Live voice runs against an ElevenLabs agent, with the session server feeding the qualification chips. Text mode replays a scripted call, so it works with no backend.

pnpm install
pnpm -r test              # qualification core, summary, eval, and server tests
pnpm evals                # run the persona eval suite

# terminal 1: session server for live voice field updates
pnpm dev:server           # http://localhost:3001

# terminal 2: Next.js UI
cd web && pnpm dev        # http://localhost:3000

Or run both together from the root: pnpm dev

Environment:

OPENAI_API_KEY is optional. Without it the summary agent uses its deterministic template. Set SUMMARY_MODEL to override the default model.
The ElevenLabs agent id ships with a default, so live voice works out of the box. Override with NEXT_PUBLIC_ELEVENLABS_AGENT_ID.

Demo shortcuts:

Live voice, autostart: http://localhost:3000/?autoplay=1
Scripted text demo: http://localhost:3000/?autoplay=1&mode=text
Text mode by hand: click "Type instead" on the home screen.

Project layout

packages/shared        Shared TypeScript contracts (events, lead, tools)
packages/core          Deterministic qualification core, tests, seed catalogue
packages/summary       Post-call summary agent, grounded against the lead record
packages/evals         Persona eval harness and rule assertions
packages/voice-client  Browser transport to the session server (HTTP + WS)
apps/server            Session API and WebSocket event bus for live voice
web                    Next.js single-screen UI (idle, call, composing, summary)
plans                  Build plans

Stack

Next.js, React, TypeScript, Tailwind
ElevenLabs for the live voice call, agent and UI
OpenAI for the post-call summary agent, behind a template fallback
Vitest for the core, summary, eval, and server suites

Built at the Fleek x a16z hackathon.

HahaBill/supconnect

README