How it works

Setting up a program

You create a program, then a cohort with a name, dates and an optional demo day. Inside the cohort you generate invitation links. Each link carries its own expiry and its own use limit — one for a single founder, or twenty for a whole intake.

Links are copied to your clipboard and shared however you already talk to your cohort. The first version sends no email. The token in the link is random and stored only as a hash, which means a link can be revoked or allowed to expire, but never recovered or re-displayed.

What a founder does

A founder opens the link, creates an account or signs in, and joins the cohort. Before anything is uploaded they are shown a short notice covering what is collected, that their microphone audio is streamed to the live-voice provider (OpenAI) during the meeting, that this service keeps no recording and no transcript, what their browser measures, what their program manager can and cannot see, and the deletion deadline. They have to accept it explicitly; the timestamp and the version of the text they accepted are recorded.

Then they supply their company information:

• A written brief — who the customer is, what the product does, traction, what they are raising and what it buys. Sensitive or unknown fields can be left blank.
• Optionally a pitch deck, as a PDF up to 4 MB. It is uploaded to private storage and its text is extracted on the server.

Either half works alone. If a deck was exported as images and carries no readable text, the founder is told that plainly and offered the deck-free path, rather than being left to assume it was read.

Choosing a meeting focus

Before starting, the founder chooses what kind of meeting to rehearse: a general investor conversation, one aimed at a firm selected from a short list, or one aimed at a firm they name themselves. General is the default, and the choice is saved with the session.

A firm selection changes the focus of the rehearsal, nothing else. It is independent practice: not affiliated with or endorsed by the firm, not a connection to it, and not a reproduction of a real investor. No firm is involved. The AI investor is not that firm and does not speak for it, and it is instructed never to say that a firm or one of its partners would ask a particular question.

It also only uses what the operator has actually researched. Each firm has a profile carrying a source, a “last checked” date and a review status, and the meeting uses a profile’s context only once a person at this service has checked it against that source. No firm supplied, reviewed or approved any of it. The context is kept deliberately small: broad stage and sector language taken from the firm’s own public website. Where none exists, the meeting is an ordinary general conversation that carries the firm’s name as its label, and the founder is told so on the setup screen before they start. The firm’s name, and any such context, is sent to OpenAI with the rest of the meeting’s material.

The meeting

The meeting is a live spoken conversation. The founder talks into their microphone; an audio-capable AI model hears them directly and replies aloud, in a synthetic voice, as a generic, sceptical practice investor. Turn-taking is automatic: OpenAI’s own voice-activity detection decides when the founder has finished speaking, and is set to wait longer when they pause or trail off, so there is no button to press at the end of an answer. Long answers are not cut off. Apart from the founder pressing End meeting, the conversation ends only at the 55-minute limit (even mid-answer), or at the response cap or the session’s usage limit described below (after the founder has had up to 60 seconds to finish their point). The founder can interrupt the investor by speaking; that works best with headphones.

It asks one question at a time and follows up on what the founder actually said. It probes vague claims, unsupported growth assumptions, customer evidence, market sizing, unit economics, competition, how they sell, gaps in the team, and whether the raise amount matches what it is supposed to buy.

Starting a session does not touch the microphone. In the meeting room, a screen first explains that audio will be streamed to OpenAI and asks the founder to switch the microphone on; the browser then shows its own permission prompt. A microphone is required — there is no typed alternative — and headphones are recommended.

The meeting itself fills the screen and is kept minimal: one animated orb in the centre shows whether the investor is listening, thinking or speaking, a quiet clock shows the time elapsed and the time left, and there are two controls, Mute and End meeting. Mute switches the founder’s microphone off, and no sound from it is sent while it is muted. There is no transcript and there are no captions.

OpenAI limits a live session to 60 minutes and offers no way to extend or resume one. Within that one session, the spoken interview has a hard maximum of 55 minutes, and up to 5 minutes are kept for the same AI session to write the report from what it heard. The investor manages the time itself — it is told how much has passed and is left, and does not cut the founder off for answering at length — and at 55 minutes the conversation ends automatically and the report is requested. The founder can end it at any time before that. Writing the report usually takes a minute or two. If it cannot be finished before the 60-minute limit, the founder is told there is no report for that meeting: an unfinished report is never saved or shown.

Time is kept by the meeting monitor, a small FAUVION service that holds OpenAI’s server-side connection to each call: it opens the meeting, keeps the AI told how much time has passed and is left (the AI mentions time only when it is useful, such as near the end), ends the conversation when time is up, and asks for the report. It receives no audio. It handles the meeting’s instructions (built from the brief and deck), the report request and the report in memory only, for the length of the call, and keeps nothing that was said after it (only, briefly, that the call closed, why, and its usage total). If it sees the session changed in a way it did not make itself, no report is written for that meeting, because one written from it could not be trusted; if it was a change only tampering could cause, the call is ended at once.

As a safeguard against runaway loops, an interview also has a cap of 200 AI responses — far more than a real meeting has. Every response the AI starts in the conversation counts: spoken replies, silent ones where it lets the founder carry on, replies cut short by the founder speaking, and its opening. Report requests never count; they have an allowance of their own. As the cap approaches, the investor is told to steer towards a close and then to close once the founder finishes their point, and the meeting screen says so (the founder can also press End meeting at any time). Once the investor has said its closing words, the AI starts no new reply; the founder can finish their point, and the conversation then ends and the report is written from the audio it heard, as usual. The last response is kept for those closing words: if the investor has not closed by then, the AI stops replying of its own accord and, once the founder has finished, is asked for its closing words. The founder is never left talking to an AI that has silently stopped: the conversation ends at most 60 seconds after the cap (plus the closing words), and the screen says why.

Each session also has a usage limit: a ceiling on what the AI provider is asked to do in it. Everything counts towards the same limit — every reply in the interview, the investor’s closing words, each request for the report, and any conversation started again in the same session after a dropped connection. The meeting monitor keeps a running total from the token counts the provider reports (counts only, never what was said), and always holds part of the limit back for the closing words and for at least two attempts at the report, so the interview cannot use up what the report needs. Most meetings never come near the limit. If one does, it ends the same way as at the response cap: the investor is told to steer towards a close and then to close, the meeting screen says the session is approaching its usage limit, the founder finishes their point, and the report is written. So a meeting lasts up to 55 minutes; an unusually costly session can end sooner. A conversation started again after a dropped connection has only what is left of the session’s limit, and a session with too little left cannot open another one — the founder is told to start a new session. If the limit is ever passed before a meeting can be wrapped up, the call is stopped and the screen says that no report could be written. As a further precaution, a call the meeting monitor has stopped confirming it is watching is ended.

There are limits per founder as well. A founder can hold one live conversation at a time, and has an allowance for any 24 hours across all their sessions; when it is used up, another conversation has to wait. Deleting sessions does not reset either. The service as a whole has an allowance for any 24 hours too, across every account; when that is used up, nobody can start a conversation until some of it is a day old.

The usage limit is a safeguard for the people running the service, worked out from the provider’s published prices rather than its invoice. Several things are done so that it cannot be got round: the browser’s connection to the provider carries audio only, so the browser has no channel on which to ask the provider for anything; the investor cannot reply at all until the meeting monitor has opened the meeting; and a call in which the monitor sees something that neither it nor the founder’s voice could have caused is ended at once, counted in full, and its session closed. It is still not described here as a guarantee: those protections rest on the provider behaving as its documentation says, which has not yet been confirmed against the live service, and the amounts are estimates.

A meeting that is left before it is ended, or whose connection drops, cannot be picked up later, because no recording or transcript of it is kept. The founder can start the conversation again — a session allows up to 3 live conversations — but each session produces one report.

Where the audio goes

The connection runs from the founder’s browser directly to the live-voice provider (OpenAI). For the length of the meeting OpenAI receives the microphone audio as it is spoken, along with the written brief, the text extracted from the deck if one was uploaded (the first 24,000 characters), and the chosen meeting focus. It also receives a pseudonymous account identifier — a one-way hash of the account ID, the same for each of the founder’s meetings, with no name or email in it — which it uses for abuse monitoring. When the meeting ends it also receives a handful of the numbers the founder’s browser measured, described below, so that the report can refer to them.

FAUVION itself never receives the audio: it is not recorded, uploaded to or stored by this service. There is no transcript either. None is shown, none is stored, and OpenAI is not asked to transcribe what the founder says. One detail is worth stating: OpenAI’s service produces a text version of the AI’s own spoken replies on the call’s event stream, and that cannot be switched off. The browser does not receive that stream — its connection carries audio only — and the meeting monitor, which does, drops the text unread.

OpenAI handles what it receives under its own terms, which FAUVION does not control. By its published API data controls, checked on 2 October 2026, API data is not used to train its models by default, and it may keep abuse-monitoring logs, which can include content, for up to 30 days, or longer where the law requires. Zero data retention and EU data residency exist, but they need OpenAI’s approval and are not arranged, so processing is in the United States by default. FAUVION cannot delete what OpenAI holds. The privacy page sets this out in full.

What the browser measures

While the meeting runs, the founder’s own browser measures the microphone signal on the device and keeps only numbers; the audio is read and dropped. It records how long the meeting lasted, how long each of the founder’s turns was and how much of the talking they did, the pauses inside their turns, how long they took to start answering, and how many times they began speaking while the investor was still speaking.

It also measures how far the founder’s pitch moved, as a range in semitones relative to their own median — absolute pitch is not kept — and, only where the browser leaves the microphone level unprocessed, how far loudness varied. Pitch and loudness figures are never sent to the AI. All of these numbers are approximate. They are stored inside the report and deleted with it, and they are measurements of sound: nothing here reads emotion, confidence or personality from a voice.

The report

The report is written once, at the end, by the same AI model that held the conversation, in its own words, from what it heard plus the brief and the deck text. The meeting monitor asks OpenAI for it over its connection to the meeting, and FAUVION’s server checks and stores the copy it receives; the founder’s browser cannot change what is stored.

On the pitch itself it covers what held up, weak or unsupported answers, contradictions and points to clarify between deck, brief and what was said, the questions to be ready for next, evidence worth preparing, and three priority actions.

Every material point names where it came from and where in it — for example, “Based on: What you said in the meeting · when asked about pricing”. The report contains no transcript and no direct quotes or excerpts of anything the founder said or wrote; nothing in it is put in quotation marks. Points that cannot be attributed are left out. There is no score and no probability of raising. Read a full example.

The report also has a spoken presentation section. It says whether each of the investor’s questions was addressed — answered, partly answered or not answered — and gives coaching on pace, pauses, clarity, filler words, structure and concision, with something to practise for each point. Those notes are the AI’s impressions from listening, not measurements, and it can mishear. Turn-taking is coached only from what the founder’s browser measured: if they repeatedly began a full answer before the investor had finished, the report gives a gentle, specific suggestion, such as letting the question finish first. A single overlap, or a brief “mm”, is not counted.

The figures the founder’s browser measured sit alongside that coaching and are kept apart from it: they are numbers about sound, calculated on the device, not by the model. Neither the coaching nor the figures are turned into statements about confidence, emotion or personality; nothing is said about accent or voice; and neither is a judgement of whether a company will raise. The figures are approximate.

The report is the founder’s. They can read it, download it as a PDF or as Markdown, or print it. Their program manager cannot open it, and cannot see or hear the meeting or the measured figures either.

What rolls up to the program

Nothing, in this version. Each report is filed under a handful of themes from a fixed list — pricing clarity, customer acquisition assumptions, unit economics, and so on — which stay with the report. The service is built to count those themes across a cohort, in fixed groups of at least 3 founders added and removed together, and show the counts to the program manager. That is switched off: independent review found that a manager could still sometimes work out one founder’s themes from the counts, and making it safe enough needs decisions about the product that have not been made. The privacy page says more.

Deletion

A founder’s deck, extracted text, brief and report — including the measured figures stored inside the report — have a 7-day deadline. It is set when they first create any of it and is never extended — coming back to the app does not buy more time, and the database enforces that as well as the application.

A deletion job is queued for that deadline as soon as there is anything to delete. The job is carried out by a cleanup task that the operator of the deployment has to schedule; once it runs after the deadline, the content is removed from the live service. Where the operator has not scheduled it yet, the founder’s own data page says so, and deleting from there is what removes the content. There is never any audio or transcript to delete: this service keeps neither. What OpenAI holds under its own terms is outside this, and FAUVION cannot delete it.

A founder can also delete everything immediately from their dashboard. That runs the same cleanup, and the status shown afterwards reflects what actually happened to both the database rows and the stored file. The privacy page covers the backup caveat, which is real and worth reading.

The demo

The demo needs no account and is a clearly labelled simulation. No AI model is involved and nothing is sent anywhere. The investor’s lines are scripted and shown on screen; they can optionally be spoken by a voice built into the visitor’s device, as a stand-in — the real product uses OpenAI’s AI voice instead.

A visitor can switch on their microphone so that their browser measures their turns for real — the audio never leaves the page — or use simulated turns. The report shown at the end is a pre-written example, not a response to what was said; like the sample report, it can be downloaded as a PDF.