Our own build · No client results claimed

A voice agent you can talk to in the browser, with its limits stated.

Recordings make every voice agent sound good. So the demo on our site is live: you press a button, allow your microphone, and talk to it. Here is exactly how it works, what it costs to run, how long it takes to answer, and where it is still weak.

This is our own demo, not a client project. Aplos AI built it over October 29 and 30, 2026. Nothing on this page is a customer result.

Talk to it now Book a free build call

Why build it live

A voice agent is hard to judge from a description

A shop owner deciding whether an AI should answer their phone needs to hear it: the voice, the pauses, what it does when it does not know. A screen recording hides all of that, and a phone demo line invites abuse and per-minute phone charges. A browser demo with hard caps avoids both.

Two modes

Ask the Aplos assistant, or call a sample shop

The first button talks to the Aplos assistant about prices and how a build works. The second plays a made-up heating and air, plumbing, electrical or landscaping shop, so a trades owner hears how their own phone would be answered. The shops are fictional and labeled that way.

No phone number

Nothing to dial, nothing to spoof

There is no public number to abuse and no per-minute carrier fee. Visitors pass a Cloudflare Turnstile check, and each gets a one-use ticket.

Stated limits

It says it is an AI

It tells you it is the virtual assistant and that the call is transcribed. We log the text of the conversation to catch problems. Your audio is not saved by us.

How it works

Four pieces, each swappable

Your microphone audio travels to a Cloudflare Worker over a WebSocket. The Worker turns it into text, asks the agent for a reply, turns the reply into speech and streams it back.

Hearing

Whisper on Cloudflare Workers AI

A simple volume-based check decides when you have stopped talking. The recording of your sentence then goes to Whisper (large-v3-turbo) on Workers AI. There is no separate speech-to-text account or key to manage.

Stack: 16 kHz audio frames over a WebSocket, energy-based pause detection, Workers AI binding.

Thinking

Llama 3.3 70B, with code that overrules it

The model answers from a fixed fact sheet about the service. Plain code runs afterward and replaces replies that claim a booking or promise a callback, which this demo cannot do. In shop mode, safety rules run without the model: if a caller mentions a gas smell, the agent tells them to leave and call 911.

Stack: Workers AI, structured JSON replies, rule-based guards, the same receptionist logic our customers' lines use.

Speaking

Kokoro, an open-source voice we host

The voice is Kokoro, an open-source model, running on our own server and played back at 24 kHz in your browser. The greeting and goodbyes are pre-rendered so they start instantly, and a short spoken acknowledgement covers the wait while the answer is prepared.

Stack: Kokoro-82M (ONNX) on a CPU server, 24 kHz PCM streaming, pre-rendered fixed lines.

Guards

Caps, one call at a time, a kill switch

Each visitor gets three sessions a day and the whole demo gets sixty. A call ends after two minutes. Only one conversation runs at a time, because the voice server is the bottleneck. An operator can switch the whole thing off, which also ends live calls within seconds.

Stack: Turnstile verification, atomic counters in D1, one-use tickets, an origin allowlist.

What we measured

Honest numbers, including the slow one

These come from scripted test calls on the deployed system, using a synthetic caller voice. They are not customer results, and we have not yet measured real callers with accents or background noise.

About 1.5 s

To the first sound after you stop

That first sound is the short acknowledgement. The greeting itself starts in under a tenth of a second.

4 to 7 s

Until the real answer starts

This is the weak spot. About two seconds go to hearing you, about two to the language model, and about a second and a half to the voice. The voice server is a CPU machine; a GPU is the planned fix.

Under a cent

Estimated cloud cost per two-minute call

An estimate from list prices for the hearing and thinking steps, not a bill. The voice runs on our own hardware.

What testing caught

Real bugs, found before launch

Stress tests found that several simultaneous calls could cut the voice stream off mid-sentence, that a stalled voice server could keep a call open past its cap, and that a malformed audio message raised an unhandled error. All were fixed and now have tests.

What we cannot claim

Not yet verified

Accuracy on real accents and noisy rooms, behavior with many simultaneous visitors, and whether the demo converts visitors into bookings. We will update this page with measured numbers once we have them.

Our own build · Open to your questions

Want this answering your phone?

Hear it first. Then decide.

Try the demo, then book a free 30-minute build call. We scope what your phone line needs and agree the price in writing before any work starts.

Book a Free Build Call