See it work | Corvair Virtual Worker

Three ways to see it

The same worker, three lenses.

Each mode shows something the others cannot. Together they cover the behaviour, the logic and the consistency.

Most convincing call

Live voice

You answer, it calls

Real time speech in both directions, in the language and voice you choose. Interrupt it. Go quiet. Answer a different question. This is where pacing, barge-in handling and turn taking either hold up or fall apart.

Book a live callarrow_forward

chat

Typed chat

Same worker, in text

The identical objectives, guardrails and outcome contract, in writing. Useful for probing the logic at your own pace, and proof that the worker is channel agnostic rather than a voice bot with a text mode bolted on.

Try a chatarrow_forward

smart_toy

Agent versus agent

Unattended, at volume

A simulated contact with its own persona and attitude, run against the worker without a human in the loop. Repeatable, which is what makes it the regression harness as well as a demonstration.

How simulation worksarrow_forward

science

A working conversational demonstrator is available today, covering live voice and typed chat across sixteen languages with eight enterprise scenarios, agent versus agent simulation, transcript analysis and a coaching pass. The broader Hub described across these pages is the platform being built around that core. We will always tell you which is which.

What to watch for

Seven behaviours that separate this from a scripted bot.

Most demonstrations are judged on whether the voice sounds natural. That is the least interesting question. These are the ones worth watching.

badgeIt establishes who it is talking to. Give it a name it was not expecting, or say you are answering on someone's behalf. It should stop, settle who you actually are, and discard what it assumed about the person it expected. It will not book a meeting with someone whose identity is unresolved.
stairsIt works the ladder in order. Offer it a meeting with yourself before it has asked about the senior person. A scripted tool takes the easy win. This one asks the higher question first, and only then accepts the alternative.
edit_noteIt takes a correction as a win. Tell it something in its briefing is wrong. It should accept the correction without defending the file, and you will see the corrected fact appear in the record at the end.
blockIt refuses to quote. Ask what this costs. It should decline cleanly, offer the account team, and not improvise a number to keep the conversation moving.
historyIt does not re-ask. Tell it you will check a calendar. It should treat that topic as settled permanently. Asking twice is how a worker earns "I already told you that".
logoutIt offers an easy exit. Say you are not happy with the call. The correct response is one sentence of acknowledgement and an offer to let you go, not a save attempt.
fact_checkIt ends with a record, not a summary. The outcome and its status, the confirmation path, every piece of intelligence gathered, and blank fields where it does not know. Read it and check whether anything in it was invented.

The last one is the real test. Ask yourself whether anything in that record is something you did not actually say. If the answer is no, everything downstream can be trusted.

Adversarial

Please try to break it.

We test against a matrix of contact attitudes precisely because a worker that performs well against a friendly contact often collapses against a busy one. Here are the ones worth trying yourself.

boltBe busyGive it thirty seconds. Does it reach a substantive ask before you leave?

sentiment_dissatisfiedBe irritatedDoes it apologise once and offer an exit, or try to save the call?

psychology_altBe scepticalDoes it offer a concrete insight, or escalate a claim it cannot support?

chat_bubbleAnswer in one wordDoes it read hesitation and slow down, or fill the silence?

person_offBe the wrong personDoes it establish your actual role, or carry on with someone else's?

forumRambleDoes it hold the objective through a tangent and return without restarting?

translateSwitch languageMid conversation. Does it follow you and keep the record consistent?

do_not_disturb_onAsk to be removedOne apology, the request recorded in your words, and a prompt end. No argument.

verified

Every one of these is a cell in the simulation matrix the worker is tested against before release, including the suppression request, which is treated as the most serious class of defect available to this platform if handled badly. You are not going to find something we have not looked for, but we would genuinely like you to try.

After the call

What you get to look at afterwards.

The conversation is only half the demonstration. The other half is what the platform did with it.

description

The transcript

Verbatim in the language spoken, with a parallel rendering in your record language, both retained.

fact_check

The outcome record

Structured, with the outcome level, status, confirmation path and every intelligence field populated only from evidence.

scoreboard

The score

What earned progress, what was left on the table, and typed findings anchored to specific moments in the transcript.

edit

The coaching pass

Three to six surgical prompt edits proposed, with the evidence attached, for a person to approve or reject.

insightsHow scoring and coaching work verified_userWhat is checked before dial paymentsWhat the call cost

Take the call yourself.