Founder · solo designer and engineer · 2026

Verified Call

Know who is on the other side of the call.

I designed Verified Call and built it with AI coding agents, a first working version in 3 weeks: a Mac and Windows app that checks the candidate, and a panel inside Google Meet that shows the interviewer what it found.

The panel inside Meet

meet.google.com

The panel layout and copy from the shipped extension, replayed live. The people are fictional portraits.

Role · Team · When
Founder, solo design and engineering · Sep–Nov 2026
Problem
The person on the call may not be the person you hire.
Key decision
Tell the host, don't block the candidate: no check waits on a click.
Outcome
3 weeks to a first working version: 5 surfaces, one design system
Stack
Electron, React, TypeScript · Chrome extension · Next.js · Swift and Python probes
30/30tries caught, for each of 4 hidden AI tools
0/3,990face swaps got through (2,855 flagged, 1,135 "Can't confirm")
20/20live Meet calls read right

01The problem

The person on the call may not be the person you hire.

Someone else sits in

A stand-in takes the interview, or a recording plays where the webcam should be.

A hidden AI reads answers

An overlay the call cannot see listens and suggests what to say.

A face swap

A live filter turns one person into another, in real time.

Why I built it

Recruiting at SigIQ, the only way to be sure a candidate wasn't cheating was to fly them in. It worked, but it was slow and costly. Verified Call brings that check to remote calls.

02The interfaces

Every screen, every state.

Real screens, rendered from the shipped code with demo data.

2.1 · Desktop app · Mac and Windows

The candidate's side: one calm flow, from invite to join.

It checks that the camera is real, that the face and voice stay the same person, and which apps are open. Then it runs quietly during the call.

Verified

HomeCalls arrive from the invite link. Each opens 5 minutes before it starts.

All other states (8)
Empty home
Call details
A quick check
What you need
Voice step
Head step
Try once more
Check ended

2.2 · Inside Google Meet

The interviewer's side: a badge on the tile, details on demand.

One click on a guest's badge shows each check, when it ran, and the fact behind any finding.

CheckingFirst results within a minute
VerifiedEach row with its time
Needs a lookAmber, with the fact named

2.3 · Chrome extension

Set up a verified call from the calendar you already use.

A switch per guest in Google Calendar, then upcoming and past calls with each guest's result.

Popup home
Upcoming
Past
In Google CalendarThe badge is the switch

03Six decisions

A product that checks people has to earn trust.

Each one shows what I tried first.

Firstv1

Alex Chen
No hidden apps Verified
Open apps VerifiedChecking
Real camera Verified
Face match Verified

ShippedFinal

01 · Calm

A result changes only when the fact does

Decision
A result stays put until a newer check changes it.
Result
Each check shows once, with its time. A "Needs a look" holds still.
Why, and what I tried first
Why
Warnings that keep firing cause fatigue, and people start ignoring them.
Tried first
Rows that redrew on every check. "Open apps" flipped Verified, Checking, Verified, and each flip rebuilt the panel.

Firstv1

Alex Chen
No hidden apps Verified
Open apps Can't confirm
Google Chrome, Slack, WhatsApp
Real camera Verified

ShippedFinal

02 · Words

Say what it saw, never blame

Decision
The panel names what it saw: which app was open and when it was used.
Result
"ChatGPT (used 9:07 PM)", and one plain next step.
Why, and what I tried first
Why
A vague "Can't confirm" read as the product failing, even when the cause was known.
Tried first
"Can't confirm" on rows where the cause was already known.

Firstv1

On, glowing green

ShippedLive

Off
03 · Colour

Green only means verified

Decision
The on/off switch is the badge itself. It turns green only when the person is verified.
Result
Graphite when off, silver when on, green only when verified. Try it.
Why, and what I tried first
Why
Green has one meaning. A switch that is merely on must not look like a pass.
Tried first
A switch that glowed green when on. I dropped it on purpose.

Blockv1

OBS is open

Please close OBS to continue.

Choicev2

OBS is open

Close OBS for the best resultContinue anyway

NoticeFinal

Ready

Your host has been told that OBS is using your camera.

Join the call
04 · Flow

Tell the host, don't block the candidate

Decision
If another app uses the camera, the candidate sees one amber note and joins. The host sees "Needs a look".
Result
No check waits on the candidate, and the host still gets the fact.
Why, and what I tried first
Why
The check is for the host. A check that waits for the candidate's click can be skipped by never clicking.
Tried first
A hard stop. Then a choice: "Close OBS for the best result" or "Continue anyway".

ShippedFinal

05 · Presence

Never join the call

Decision
The check runs on the candidate's computer. Nothing joins the meeting.
Result
Two people in the call, and a badge on the guest's tile.
Why, and the trade-off
Why
No bot in the participant list. The call looks and feels like any other call.
Trade-off
The candidate installs a small app once, and the result shows on the host's own screen.

Black tabv2

ShippedFinal

06 · System

Black only for the main action

Decision
The extension uses the light palette. Black is kept for the one main button.
Result
Light panels, and the selected tab uses the same glass button as every surface.
Why, and what I tried first
Why
A heavy all-black popup did not match the rest of the product.
Tried first
An all-black popup, then a flat black tab.

04Evidence

How I know it works, and where it doesn't.

A detector graded by the person who built it will flatter itself. So I set up the evaluation before writing the checks, and I publish where they fail.

4.1 · The evaluation system

See the evaluation code ↗
  1. Pinned

    Rules before results

    Read more

    Every pass/fail rule is written down and fingerprinted (sha256) before a test runs. Once pinned it can't change; a new idea means a new version, tested again.

  2. Sealed

    Sealed test data

    Read more

    Development, sealed and held-out sets are split by a salted rule. Checks are built on development data only; nobody tunes against the sealed set.

  3. Graded apart

    A separate verifier

    Read more

    A fresh agent with no build context is the only one allowed to open sealed data and mark a result. The builder can't grade their own work.

  4. Bounded

    Honest statistics

    Read more

    Results carry exact 95% lower bounds. 29 tries per attack family before claiming 90%. Failed and invalid runs stay on record, like a camera rule that scored 14 of 17 before it was rewritten.

  5. Copy-checked

    The copy is tested too

    Read more

    A check fails the build if the product's own words claim more than the tests support, such as "verified human" or "deepfake-proof".

  6. Re-attacked

    One command to attack it

    Read more

    The red team is pre-registered and re-runs with a single command, so every fix is checked against every known attack.

4.2 · Results

What it measured

Hidden AI tools

30/30

Each of 4 tools (Interview Coder, OpenCluely, Pluely, Cheating Daddy) found in 30 of 30 tries; 0 false alarms in 60 clean runs. Own testing, October 4, 2026.

Camera software

29/29

29 of 29 tries for each of 9 kinds of camera software and 7 direct cameras (September 30); 20 of 20 live Google Meet calls read right (October 2).

Face swaps

0/3,990got through

0 of 3,990 swap sessions got through: 2,855 flagged, 1,135 ended "Can't confirm". The real person was cleared in 962 of 970. Recordings of 70 people, October 1.

Speed and cost

4.66s · p95, Mac

A switch to camera software in the middle of a call is caught within 4.66 s on Mac and 5.34 s on Windows (95th percentile). The camera check uses 1.27% and 0.57% of the CPU.

Records that can't be forged

80/80

Every result from the candidate's computer is signed and chained. 80 of 80 forged or altered records were rejected; 4 of 4 genuine sessions accepted.

4.3 · Limits and customers

Where it doesn't work yet

I attacked my own system 17 ways (October 17). Six still get through:

  • a capture card that copies a real camera's full identity
  • video relayed as a built-in camera
  • a second computer beside the first
  • joining with the camera off
  • a copied Windows signing key
  • a renamed helper app that doesn't hide

Linux has a known camera spoof, and voice is measured offline only, not in live calls.

What early conversations told me

20companies, closed beta
5hiring teams talked to
0live interviews yet

The closed beta is open to 20 companies. I've talked with 5 hiring teams; no live interviews have run through it yet. Their reaction was consistent: if it works as tested, they'd use it.

The one change a customer made

Asked
My first customer conversation asked to see every app open on the candidate's computer.
My line
That collides with the candidate's privacy, so I drew a line: the interviewer always sees an app used during the call, and the full list only if the candidate agrees.
Next
The first live interviews, and a field false-alarm rate published beside the lab one.

05Fair to the candidate

A finding is advice, not a verdict.

Verified Call decides what an interviewer sees about a person who may be hired or not. These are the rules that keep it fair to the person being checked.

  1. What "Verified" means

    Read more

    The camera is a real camera, the face and voice in the call match the same person's check minutes earlier, and no hidden AI tool was found on their computer. It does not check an ID, and it can't see a phone or a person off camera.

  2. Events, not accusations

    Read more

    The panel says what it saw and when: "used during this call", or "open in the background" if an app was only open. Camera and face findings carry two stills as proof. Note-taking apps are listed, not flagged. What it means is the interviewer's call.

  3. The host is told; the candidate isn't blocked

    Read more

    When a hidden AI tool is used, the interviewer sees "Needs a look" at once, with a suggested next step such as asking the candidate to close other apps. The product itself never stops the candidate or tells them to close anything: what to do is the interviewer's call.

  4. When a check can't run

    Read more

    A dropped connection or a check that can't start shows "Can't confirm": not done, not failed. The candidate can try again, and the other checks keep running.

  5. The candidate can say no

    Read more

    A consent screen comes first. They can decline and still join; the interviewer sees only that the check wasn't done. Their list of open apps reaches the interviewer only if they agree.

  6. Kept for days, not forever

    Read more

    No recording of the screen or the call. Face measurements are deleted when the call ends; check results after 7 days, or sooner on request.

  7. Open problems

    There is no way yet to contest a finding inside the product, and no written policy for assistive technology such as dictation software. The design I want: the candidate sees the same fact the interviewer saw, and can add one line of context that sits beside it.

06How it is built

An app, an extension and a server, on one design system.

Desktop appVerified serverChrome extension CANDIDATE · MAC, WINDOWSNEXT.JS WEB APPINTERVIEWER · CALENDAR, MEET Electron and ReactReal camera checkOpen and hidden appsFace (MediaPipe), voice Invite links and statusFace match (Rekognition)Speech (Deepgram)One result per guest Switch per guestBadge on each Meet tilePanel with time and proofReads the status check resultsstatus per guestinvite, switch on A verified:// link in the candidate's invite opens the desktop app on the right call. Desktop appVerified serverChrome extension CANDIDATE · MAC, WINDOWSNEXT.JS WEB APPINTERVIEWER · CALENDAR, MEET Electron and ReactCamera, apps, face, voiceFace match, speech, statusOne result per guestBadge on each Meet tilePanel with time and proof check resultsstatus per guest
Show how it works

From the code · the camera notice

// QA-55: "Your host has been told that Python is using your camera."
// Each text is notified once, until it changes or goes away.
export function cameraNoticeToShow(st: CameraNoticeState, s: ConditionsSnapshot | null): string | null {
  const text = cameraNoticeText(s);
  if (text === st.shown) return null;
  st.shown = text;
  return text;
}

The candidate is told once, and told again only when the fact changes. The same rule runs the host's panel.

When one part fails, the rest keeps going"The face check couldn't start. The rest of the check keeps running."
Connection drops are calmThe app says it is reconnecting and asks the candidate to keep it open. The check carries on.
Nothing waits on a clickEvery check starts on its own. A busy camera retries, then gives the host a plain line.

Mine

All product design, the brand and the design system. All the engineering: the Mac and Windows app, the Chrome extension, the server and web app, and the website.

Borrowed

MediaPipe for face landmarks, Amazon Rekognition for face match, Deepgram for speech to text, GSAP for site motion, and two open fonts, Turret Road and Manrope.

AI as a tool

AI coding agents wrote most of the code; I directed it. I wrote the specs, the evaluation rules and every product and design decision (about 100 of them, logged with dates), reviewed each change, tested it on real machines, and caught what the agents got wrong. A separate fresh agent checked each result before I accepted it.

Engineering

Two hard problems.

01 · Desktop app

Catching camera software without trusting its name

Problem
A face swap has to enter the call somewhere, and on a laptop that is almost always camera software: a virtual camera between the swap tool and the meeting. Names don't help, because a virtual camera can call itself "FaceTime HD Camera".
Show how it works
Approach
So the check never reads names. A small native probe (Swift on Mac, Python on Windows and Linux) asks the operating system what actually backs each camera: a signed camera extension, a legacy plug-in, a virtual transport, or a USB or built-in device. It works out which apps are reading each camera from the capture stack's own log lines, because a virtual camera that is only running proves nothing. A pure, version-pinned rule then gives one of three answers, and "unresolved" is a real answer, not a guess.
Result
It took one failed rule version (14 of 17) and a 178-trial name-scramble test to get there; the sealed result was 29 of 29 per attack family on Mac and Windows, at about 1% of one CPU core.
From the code · the camera ruleTypeScript
function classify(d: ProbeDevice): { cls: DeviceClass; reason: string } {
  const kind = d.backing?.kind ?? "unknown";
  if (kind === "camera-extension") return { cls: "software", reason: "device-backed-by-camera-extension" };
  if (kind === "dal-plugin") return { cls: "software", reason: "device-backed-by-dal-plugin" };
  if (d.transport === "virtual") return { cls: "software", reason: "device-transport-virtual" };
  if (PHYSICAL_BACKING.has(kind)) return { cls: "physical", reason: `device-backed-by-${kind}` };
  return { cls: "unknown", reason: `device-backing-${kind}` };
}

02 · Chrome extension

A Google Meet panel that survives Meet's redesigns

Problem
Building inside someone else's app means building on ground that moves. Meet's class names are generated and change without notice, so the extension never reads them.
Show how it works
Approach
It finds the control bar the way a person does: the leave button, by its accessible name in 8 languages or by the icon Meet draws the same way in every language, then the nearest group of buttons around it. Placement is idempotent and self-healing: put back when Meet re-renders, moved when the layout changes, and a floating badge if Meet changes beyond recognition, never a silent failure.
Result
Saved Meet page fixtures (2, 4 and 8 tiles, pinned, presenting, camera off) turn a Meet redesign into a red test before an interviewer ever sees it.
From the code · finding Meet's control barCode comment
// Meet's class names and jsname values are generated and change between releases, so this file never reads them. It
// finds the bar the way a person would: the leave-call button (its accessible name, or Meet's "call_end" icon, which
// does not change with the page language), visible, in the lower half of the window and not inside a dialog or menu.

07Versions and dead ends

What I tried before the final.

Website

Why the final won: it reads in five seconds, with one headline, one board and one button.

Bento waitlist v1
Three-column rails v2
Hero with film v3
Built to stop Final
Guide page v1
Guide template Final
Comparison
Candidate trust
Split-flap 404

Hero and research

Why the final won: still concepts first, and the winner shows the real check, not a face.

Concept A: flow
Concept B: record
Concept C: two views
Research v1
Research Final

Extension and Meet panel

Why the final won: a light popup with one black action, and a panel that names the fact.

Home v1
Home Final
Glass tab v3
Panel v2
Calm panel Final

Badge, icons and motion

Why the final won: one badge and one pixel language, so every surface reads as one product.

Shield badge lab
Materials and sizes
Pixel drawing board
Animated stat icons
Motion rules
Design system 1.3.0

08Brand system

The brand system

Split-flap · the signature motion
Verified
V
One tile, slowed down
Badge, pill and button
The Verified badge
Verifying2 / 2 verified1 needs a look
Join the call
Display · Turret Road
Aa

0123456789

Text · Manrope
Aa

Plain words, sentence case

Colour · one means verified
Paper
Ink
Hairline
Tile
Verified
Caution
Pixel icons · 20 × 20

09What I learned

Three rules I keep.

Still concepts before motion

The particle-face hero read as uncanny. Now I show two or three still options and pick one before anything moves.

Calm is a feature

The panel got better each time it did less. A signal that stays still is a signal people trust.

No check may wait on a click

A choice screen once held the face check until the candidate pressed a button. A check that can be skipped by not clicking is not a check.

I design and build the whole product.

If your team needs a design engineer who ships the interface and the code behind it, I would like to hear from you.

Designed and built by Bhanu Kowshik · LinkedIn ↗