The assistant that disappears when you share your screen.
A desktop AI that listens to your live technical sessions, answers in real time, and stays excluded from screen capture. Prepare on the web. Perform live. Pay only for what you use.
Problem
AI help stops exactly where the pressure starts.
Interviews, technical screens, client demos, architecture reviews. The moments that decide outcomes now happen in real time on Zoom, Meet, and Teams — with no pause button and no second window to think in.
Alt-tab to a chat window, type the question, read, come back. The context switch costs you the thread of the conversation, and it is visible in your eyes, your cursor, and your silence.
The one moment you most want a second brain is the moment your screen is the thing being watched. So people don't reach for it at all.
People spend weeks preparing with AI, then walk into the session that actually counts with nothing.
Why now
Three things became true at the same time.
The shared screen became the desk
Remote-first hiring and remote-first engineering made screen share the default surface for high-stakes technical work — not a special occasion.
Capture APIs finally cooperate
Per-window capture exclusion, screen-share detection, and system-audio loopback are now dependable on both macOS and Windows. Genuine invisibility is an OS feature we can build on, not a hack.
Real-time multimodal got cheap
A screenshot analysed and answered in seconds costs cents. That is the difference between an impressive demo and a business with a gross margin.
None of the three was true three years ago. All three are table stakes for what we've already shipped.
Product
One account. Two surfaces. Prepare, then perform.
Where the practice happens
- Voice-driven AI mock interviews with role-specific questions and written feedback
- Skill gap analysis across algorithms, system design, and behavioural rounds
- Resume generator and job-description match scoring
- Wallet, billing, and account management
Where the session happens
- Live Captions — continuous transcription of meetings and calls
- Live Assist — question detection and real-time answers
- Visual Assist — screenshot a problem, get analysis and a fix
- Coding mode with OCR, language detection, and diagram rendering
The two halves share one wallet, one set of models, and one context. Documents you upload while preparing are the documents the live assistant answers from.
How it works
Audio in, answer on screen — without ever being on the shared screen.
Capture
Microphone, or system-audio loopback so the other side is captured cleanly through your headphones.
Transcribe
Streaming speech-to-text produces timestamped final utterances continuously.
Filter
On-device voice identity check drops your own speech so your answers never loop back in as new questions.
Detect
Pattern recognition decides whether an utterance is actually a question worth answering.
Answer
Hosted GPT or Claude answers with your uploaded documents and session context. No user API keys.
Surface
The answer appears in a transparent overlay that screen capture cannot see and the taskbar never lists.
Every step runs on the user's machine except the model call — which is why the round trip stays inside a conversational pause.
Defensibility
The invisibility is not a CSS trick.
The window is marked protected from screen capture, hides itself the instant screen sharing begins, and restores when it ends. It never appears in the dock or taskbar, and mouse events pass through it so it can't steal a click at the wrong moment.
In microphone mode the other person's voice comes out of your speakers and back into your mic — and so does yours. We enrol about twelve seconds of your voice locally, build a 39-dimension MFCC embedding plus a median-pitch gate, and drop your own utterances before they ever reach a model. No cloud speaker-ID service, no audio leaving the machine for identification, and a per-user threshold calibrated against held-out samples of your own speech.
A timestamped rolling audio journal maps every transcript result back to the exact audio that produced it, so identification runs on one sentence rather than a blend. Getting this wrong is why naive versions of this feature leak.
Notarized macOS builds, signed Windows installers, an auto-update pipeline, encrypted local configuration, and hosted models so no user ever pastes an API key. Distribution on desktop is a months-long tax that a web competitor has not paid.
Interface
Operated blind. No mouse, no window management, no tell.
Every action is a chord you can hit without looking away from the person you're talking to.
Windows uses the same chords with Ctrl. The overlay is semi-transparent and click-through by default, so the app you're actually being watched in never loses focus.
Positioning & category risk
We'd rather raise the hard question than be asked it.
This category attracts a fair question: is an invisible assistant a cheating tool? Our answer is in where the product is pointed and where users actually spend time.
- Where it's pointed
- Everyday technical work — debugging sessions, client calls, architecture reviews, standups, meeting transcription — plus interview preparation, which is what the entire web product is.
- How the funnel is built
- Mock interviews, skill gap analysis, and resume tooling are the front door. The majority of user time is designed to be practice, not live assist.
- The honest risk
- Employer and proctoring-platform pushback on live assistance in assessed settings. It is a real risk and we underwrite it rather than argue it away.
- The hedge
- A sanctioned team and enterprise mode — assistant on the record, admin-controlled, logged — where invisibility to capture is a privacy feature rather than a concealment one.
Keep this in the deck
Investors in this space will ask about regulatory, platform, and reputational exposure within the first ten minutes. Naming it first converts an objection into evidence that you have thought about the business, not just the demo.
If you have counsel's view, a platform policy read, or an enterprise design partner, this is where it goes.
Business model
Pay-as-you-go, priced off a real cost floor.
$0.05/min
Live transcription, billed by the minute of session.
≈$0.23
Typical assist at roughly 2K input and 1K output tokens.
$5 · 10 · 25
Minimum top-up $5. Razorpay, PayPal, and Stripe wired up.
None
Free signup, no card, no monthly commitment, no API keys.
A worked session. Forty-five minutes of live assist with twelve coding assists is about $2.25 of transcription plus about $2.76 of model calls — roughly $5 of wallet, or exactly one credit pack. The smallest purchase we sell is one real session.
Why metering beats subscription here. Demand is spiky: an interview loop is three weeks, not three years. A subscription gets cancelled the day the offer lands. A wallet balance doesn't churn — it waits, and it's still there the next time the user needs us.
Fill before sending. Gross margin — the spread between this rate card and your actual provider cost — is the number investors will go to first. Pull the real blended figure from your billing data and put it on this slide.
Market
Sized bottom-up from our own rate card, not from an analyst chart.
Every input below is a lever you can defend in the room. Replace the bracketed values with your own and the arithmetic carries itself.
| Input | Value | Where it comes from |
|---|---|---|
| Reachable users | [ fill ] | Technical professionals in remote-first roles who screen-share weekly |
| Signup → funded wallet | [ fill ]% | Your actual funnel conversion from the last 90 days |
| Sessions per funded user / year | [ fill ] | Billing data — interview loops plus recurring work sessions |
| Revenue per session | ≈ $5.00 | Derived: 45 min transcription + 12 assists at published rates |
| Annual revenue per funded user | sessions × $5.00 | Product of the two rows above |
Only the $5.00 session figure is derived from published InvisibleMe pricing. Every other row is yours to supply — never present bracketed values to an investor.
A defensible bottom-up model with three honest assumptions beats a $40B top-down TAM slide that anyone can puncture in one question.
Traction
What's shipped, and what still needs a source.
Verifiable today
Shipped on both desktops
Signed and notarized macOS builds and signed Windows installers, with an auto-update channel in production.
Payments and wallet live
Razorpay, PayPal, and Stripe integrated against a metered wallet with per-token and per-minute accounting.
Hosted model infrastructure
GPT and Claude served through our own backend — users never supply keys, and we control routing, cost, and margin.
Answer-engine surface
Structured entity facts, llms.txt and llms-full.txt, and intent-matched landing pages already live on the marketing site.
Currently claimed on the website
12,000+
93%
85%
4.8
These four are marketing-site figures, not billing-verified metrics. Investors diligence exactly these numbers. Before this deck leaves your hands, either back each one with analytics and billing data or replace the slide with the metrics you can defend: funded wallets, monthly top-up volume, sessions per week, and 30-day retention.
Competition
Others do one half. We're the only one holding both.
| Player | Hidden from capture | Real-time answers | Prep + live, one account | Own voice filtered | No subscription |
|---|---|---|---|---|---|
| InvisibleMe | ● | ● | ● | ● | ● |
| Live-assist rivals | ● | ● | ◐ | ◐ | ○ |
| Interview-prep platforms | ○ | ○ | ◐ | ○ | ○ |
| Meeting notetakers | ○ | ◐ | ○ | ● | ○ |
| Raw ChatGPT / Claude desktop | ○ | ◐ | ○ | ○ | ○ |
● full · ◐ partial · ○ none. Competitor rows are generalised by category and are not verified vendor-by-vendor. Name the specific companies and re-check every cell before showing this — one wrong claim about a named competitor costs you the room.
The wedge isn't invisibility on its own — it's that the same account, wallet, and document context carries a user from practice into the live session. Prep products can't follow you into the call; live tools have nothing to sell you before it.
Go to market
Distribution is a search problem, and the machine is already built.
Answer-engine optimisation
The site already ships llms.txt, llms-full.txt, and structured entity facts — a cite-ready description of the product written for the assistants people now ask instead of searching. When someone asks an AI how to get help in a technical session, we are built to be the cited answer.
Intent-matched landing pages
Dedicated pages for interview helper, technical interview assistant, coding interview help, and interview support — each answering one query directly rather than routing everything through a homepage.
Product-led entry
Free signup, no card, no API key. The first useful assist costs a user cents, and the smallest pack we sell buys one complete session.
Communities, campuses, teams
Developer communities and university career centres for the prep funnel; a sanctioned team tier to open the seat-based revenue line that metered usage alone can't reach.
Add numbers. Put current CAC, organic traffic, and signup-to-wallet conversion on this slide. GTM slides without a funnel number read as a plan rather than a channel.
Roadmap
Each phase removes the thing blocking the next one.
Stronger speaker separation
Replace the hand-built MFCC embedding with an ONNX speaker model so the filter holds up when both voices sit in the same pitch range — the one known weak point in the current design.
Team and enterprise mode
Sanctioned, logged, admin-controlled deployments. Opens seat revenue and converts the category risk on slide eight into a governed product.
Context-aware integrations
Awareness of the surface the user is actually working in — the conferencing tool and the coding environment — so assists arrive already knowing the problem on screen.
On-device fallback
A small local model for captions and simple assists: zero latency, zero marginal cost, and a free tier that costs us nothing to run.
Add dates. Attach quarters and the headcount each phase assumes, so the roadmap and the ask on the last slide reconcile.
Team
Who built it, and who you're hiring next.
- Founder
- Name, role, and the one credential that makes this the obvious team for this problem.
- Why us
- Prior experience in audio, desktop, or ML that explains why the hard parts on slide six were tractable.
- First hires
- The two or three roles this round funds, and what each one unblocks on the roadmap.
Every line above is a placeholder. Fill all three.
A signed and notarized cross-platform desktop app, a metered payments backend with three providers, a real-time audio pipeline with on-device speaker identification, hosted multi-model routing, and a marketing site engineered for answer engines.
Say plainly how large the team that shipped all of that is. If it's small, that is the strongest sentence on this slide — say it out loud rather than leaving it to be inferred.
The ask
What we're raising, and what it buys.
Amount
Instrument, target close, and any committed portion.
Months
What this buys you, and the milestone it reaches.
Split
Engineering, go-to-market, and infrastructure as percentages.
Name the milestone, not the runway. The strongest version of this slide reads: this round takes us from [today's funded wallets] to [target], ships team mode, and gets us to [revenue] — which is what makes the next round obvious.