The product
Everything MIRA does, and the promise attached to each of it.
MIRA is a web platform for planning, authorising and reviewing disclosed AI delegation. It is currently a private-beta product, not a generally available service; live meeting and AI providers remain release-gated, and numbers below are release targets unless explicitly labelled measured.
Before the features
Four things MIRA will never be, whatever anyone offers us.
Half of this product is a refusal. Saying the refusals out loud, first, is how they survive contact with a roadmap.
Decision-maker
No commitments, offers, acceptances, pricing confirmations, or contract language — ever. It is enforced outside the model, so it isn't a setting anyone can relax on a Friday afternoon.
Covert impersonator
MIRA always discloses. There is no stealth mode to buy, unlock, or ask for. Undisclosed operation breaks our terms, and disclosure is checked automatically on every single meeting.
Deepfake video mask
Voice-first, with a branded visual card. A cloned voice is opt-in behind verified consent; photoreal video avatars are deliberately not on the roadmap — we'd rather lose the customer than win that way.
Notetaker or chatbot
Transcription tools capture the meeting you still had to attend. MIRA replaces the attending. And it answers from what it knows about you — asked anything else, it says so and logs the question instead of improvising your business.
The mirror
It listens to exactly one person in the room.
Your vocabulary, your signature questions, the way you evaluate, your tone with different audiences, and your boundaries. Everyone else in your meetings is context — never voice, never a fact about you.
planned
Link Gmail, Slack, your Zoom and Meet recordings, and your calendar. Each source connects — or disconnects — on its own.
Our promise Connecting or revoking a source takes under two minutes, permissions stay read-only wherever the platform allows it, and revoking one stops ingestion within fifteen minutes.
prototype
MIRA models your speech alone. Other participants inform the conversation, but nothing they say is ever attributed to you or folded into your profile.
Our promise Per-speaker audio separation wherever the platform provides it, voice matching against your enrolled voiceprint everywhere else, and independent audits of what ended up in your profile.
prototype
A 25-minute cold-start interview with MIRA, plus a handful of things you've already written — bios, past proposals, your FAQ — seeds the profile before any history exists.
Our promise A new owner can send a usable proxy within 48 hours of signing up. No six-week training period, no watching it get you wrong in front of people who matter.
planned
A nightly pass turns the day's signals into an updated profile. When something you said last month conflicts with something you said today, today wins — and last month is marked superseded rather than quietly erased.
Our promise You get a “what I learned about you today” digest each morning, and every fact shows you its source, its date, and how sure MIRA is about it.
coming next
Learning from email and Slack uses only messages you wrote. Other people's replies build context; they never build your voice.
Our promise Every learned fact is tagged with the thing you authored that produced it, and automated checks look weekly for anything that leaked in from someone else.
The live proxy
Three things have to be true at once, for twenty-five minutes.
The latency has to disappear, the disclosure has to feel human rather than legal, and the boundary has to hold against someone who genuinely wants a yes today. Any one of them failing ruins the other two.
planned pilot
MIRA joins Zoom, Meet, or Teams as a named participant with a branded card. You dispatch it from your calendar or dashboard in two clicks.
Our promise It shows up. Successful joins run above 99% of dispatches, it's audible within ten seconds of admission, and it waits politely in waiting rooms. A silent no-show pages us, not you.
simulated
Sub-second turn-taking, no dead air, and it stops mid-sentence when you interrupt it — the way a person does.
Our promise Voice-to-voice under 800ms for half of all turns and under 1.5s for 95% of them; interruptions land within a quarter of a second. Measured on every meeting, not in a lab.
prototype
A pitch, a screen, a parent sync, and an internal check-in are four different registers. MIRA picks yours from the calendar and the first minute, and you can pin it whenever you'd rather choose.
Our promise The register is visible before the meeting and overridable at any time. If it picks wrong, that's one click, not a support ticket.
prototype
It says it's a proxy inside the first two minutes, or before the first real question — whichever comes first — in your register, warmly, never as a robotic preamble. If someone new joins, it says it again.
Release target Advance notice and spoken disclosure before substantive conversation, checked automatically in 100% of sessions. No counterpart study has been run yet.
prototype
It cannot accept a deal, confirm a price, agree terms, extend or accept an offer, or book time you haven't pre-authorised. Under pressure it deflects gracefully and logs the ask.
Our promise Zero commitments, enforced outside the language model rather than requested in a prompt. Every refusal becomes a decision waiting for you. How it's enforced →
prototype
Asked something it doesn't know about you, it says so in character, captures the question word for word, and hands it back.
Our promise It will not invent your prices, your positions, or your biography. When we get this wrong we treat it as a defect and publish it on the failure-modes page.
coming next
Watch the transcript as it happens, whisper a correction mid-meeting, or take over — MIRA exits with a natural handoff line rather than vanishing.
Our promise A whisper reaches the conversation within ten seconds; a takeover completes within fifteen, with a clean verbal handover so nobody is left confused.
coming next
The default voice is distinctly MIRA's and clearly not yours. A cloned voice is opt-in, behind a recorded consent phrase and an identity check.
Our promise The clone is inseparable from disclosure and the audio is watermarked. There is no configuration in which your voice speaks without saying it's a proxy.
coming next
Track up to six people, use their names, and keep who-said-what straight in the digest.
Our promise Attribution above 95% in four-party meetings — and anything below our confidence floor is flagged in the digest rather than guessed.
After the meeting
A memory you can edit, and a mirror you can smash.
You see every belief the system holds about you and where it came from, and you can delete any of it — up to and including all of it, permanently, in one action.
prototype
TL;DR, key points, red flags, what they asked for, the decisions waiting on you, and a suggested reply — readable in ninety seconds.
Release target In your inbox within five minutes of the meeting ending, with evidence-linked actions. Current persistence and in-app notification are a backend foundation; delivery latency and the complete action set are unverified.
prototype
Every “I'll check with her” becomes a tracked item with the verbatim ask, the timestamp, and the context attached.
Our promise It is the only channel through which a decision can happen. Nothing your proxy declined gets lost between the meeting and you.
prototype
Browse everything MIRA believes about you — facts, heuristics, tone registers — each with a “why I think this”. Correct or delete anything.
Our promise Edits take effect within sixty seconds, so a correction lands before your next meeting rather than after your next embarrassment.
UI prototype
One action erases your profile, transcripts, audio, and everything derived from them.
Our promise Your key is destroyed immediately, so no MIRA system can read your data again; hard deletion within 24 hours, backups within 35 days, and a deletion certificate for your records. The full path →
prototype
Per meeting type: what MIRA may discuss, which ranges it may mention as “typically”, and which subjects force an immediate handback. Plus your public booking link.
Our promise Scope rules are enforced by the same mechanism as the boundaries, not by a prompt. Anyone booking through your link sees the proxy notice and the recording notice before they accept.
coming next
The person who met your proxy gets a courtesy summary of what was captured and a one-tap “did this feel fair?”
Our promise Their answers are published quarterly, whether or not they flatter us. This quarter's numbers →
One engine
A hiring screen and a parent sync are the same product wearing different clothes.
Listening, register-switching, boundary enforcement, and summarising are the engine. Everything that makes a conversation a screen or a discovery call is a pack that mounts on top — question banks, evaluation rubrics, register presets, boundary policies, digest templates, and the right consent text for where you are.
A pack can make MIRA stricter. Nothing in a pack can make it looser — boundaries only ever tighten, whichever pack you install.
What we publish
Numbers you can hold us to, not adjectives you have to believe.
Two of these have no acceptable drift: disclosure and commitments. A miss on either is an incident with a public write-up, not a line item in a backlog.
| What we promise | How it's measured | Where it stands |
|---|---|---|
| Every meeting disclosed | Automated transcript and playout evidence. | Release target: 100%; live evidence pending |
| Never a commitment | Adversarial suite plus exact playout reconciliation. | Release target: zero; full corpus and production evidence pending |
| Conversation stays natural | Instrumented mouth-to-ear latency under representative network/audio conditions. | Target: P50 ≤800ms, P95 ≤1.5s; unmeasured |
| It shows up | Successful joins ÷ dispatches on each certified provider. | Target: >99%; no provider certified |
| Your digest arrives fast | Meeting end to successful delivery. | Target: P95 ≤5 minutes; unmeasured |
| Corrections take effect | Accepted profile edit to authorised retrieval. | Target: ≤60 seconds; backend pending |
| It represents your authorised context | Blinded comparison on a per-owner question set with citation checks. | Evaluation design in progress |
| The other side is heard | Counterpart comprehension and fairness study. | Target: ≥4.2/5 and ≥90% of matched human baseline; study pending |
| Erasure is real | Key destruction, provider purge, restore-after-delete and backup expiry drills. | Design target: unreadable immediately, primary purge ≤24h, backup expiry ≤35d; unverified |
The standard we hold
An invisible proxy that silently no-shows a meeting is a fireable offence. Ours.
Underneath
Four moving parts, and why each one is shaped the way it is.
It joins as a guest, not a plug-in
A meeting layer across Zoom, Meet, and Teams with per-speaker audio, held behind our own abstraction so the way it joins can change when platform rules do — without changing anything you'd notice. Marketplace listings, so it's a recognised participant rather than a tolerated bot.
Text in the middle, on purpose
Speech to text, then a model conditioned on your profile, then speech again. Slower on paper than a direct speech-to-speech model — and chosen anyway, because the boundary check needs a sentence it can veto before it becomes sound.
Facts with expiry dates
Your profile is a set of dated, sourced, editable facts rather than a blob of fine-tuning. That's what makes “why do you think that?” answerable, and what makes deleting something actually delete it.
Outside the model, deliberately
Disclosure and boundary enforcement run as separate deterministic components on every outbound sentence, with your data under a key only you can destroy. Safety that lives in a prompt is safety that argues back.
More on how the boundaries are enforced, what happens to your data, and what currently goes wrong: Trust & safety →
Hold us to it
The promises above are worth exactly what you can hold us to.
Tell us which of them matter for the meetings you actually sit in — and which ones you'd need to see proved before you'd put a proxy in a room.