Proof of Concept

Premium voice for Vantage — with Claude as the brain

Claude brain · premium WebRTC voice on Cloudflare · July 18, 2026
Eight cards, ~30 seconds each. What we're proving, how we'll test it, and what it costs. Full proposal downloadable below.
EDGEPOINTE · VANTAGE
What we're proving

Three things at once, in one small test PoC

We move the AI Receptionist and Ask Vantage onto one voice stack that runs on Cloudflare.

  • Off OpenAI voice — replaced with premium ElevenLabs voices: a Coral-style female matching the Coordinator voice Hidden Creek uses today, plus a warm Ash-style male.
  • Claude becomes the brain, replacing the smaller Cloudflare model.
  • More real automation — booking, quoting, lead capture that actually fires reliably.
Tested on our own site — the EdgePointe Vantage tenant. Our property, zero client risk.
Architecture

Everything stays on Cloudflare

WebRTC audio
Cloudflare Realtime SFU — global, at the edge
Orchestration
Durable Object voice agent — memory, tools, per-tenant isolation
Listening
Deepgram speech-to-text, running in-network
Thinking
Claude via Cloudflare AI Gateway
Speaking
ElevenLabs — the chosen female & male voices
One platform. Per-tenant isolation, no in-network egress, flat cost — the model our pricing depends on.
Control

A switch only EdgePointe can flip Staff-only

Provider selection lives in the staff Cockpit — never in the customer console. Owners have no visibility into which brain or voice they run. Every change is written to the audit log.

Brain
Claude (primary) → Cloudflare AI (fallback)
Voice
ElevenLabs (primary) → OpenAI TTS (fallback)
Automatic and manual failover — and a written plan to retire the fallbacks once the new stack proves itself.
Training & guardrails

Your training carries over — and the guardrail gets stronger Non-negotiable

  • Nothing to redo. Claude reads the exact same owner training notes, persona, and website content you already maintain. The training screen doesn't change.
  • Closed-book. Both assistants answer only from the website and documents — anything else gets a safe "let me take a message," never a guess.
  • Enforced by tests — zero fabricated prices, zero unsupported claims.
Bonus: Claude's larger memory lets us feed in far more of the site and real documents than we can today.
Cost

Cheaper than what we're leaving Favorable

Proposed stack
~$0.05–0.06 per minute — about 20 cents on a typical call
Today's OpenAI voice
~$0.15–0.30 per minute
Why
The brain is a small slice of a call; premium voice is the main cost — and we control it
Better brain, better voice, lower cost. Estimates the PoC will confirm with real per-call numbers.
How we test it

Scripted calls, measured head-to-head

One tenant, behind a flag that's off by default. Scripted calls in the browser — no phone number needed, so it proceeds ahead of the EIN.

  • Book a tour · ask hours · ask price (quote the rate card only) · capture a lead · one Spanish turn · take a message · interrupt mid-sentence.
  • Measured against the current stack on latency, tool accuracy, voice quality, and real cost.
Go / no-go: under 800ms voice-to-voice · 95%+ tool accuracy · zero fabricated prices · cost at or below today.
Status

Approved — and underway

The proposal and build plan are approved; scaffolding has begun.

Needed to start
ElevenLabs key + the two chosen voice IDs, sign-off on the Ash-style male, and two Cloudflare dashboard clicks
Timeline
Roughly one to two weeks to a go/no-go readout
Risk
Flag off by default — production is untouched throughout
Download the full proposal below for the architecture, cost model, and success criteria. — Paul
Card 1 / 8 · ~20s