Problem Architecture Geometry Measured Boundaries Try it yourself Private demo Security & roadmap Request pilot →
For remote hiring  ·  Patent pending

The person you interviewis the person you hire.

Verified. Not remembered.

Verlinko anchors a candidate's face at the first interview, then confirms it at every round, on Day One, and at unannounced moments through their tenure. A paired second camera makes deepfakes and hidden AI help physically hard to pull off. No recordings, no documents, no names stored.

Runs alongside Zoom, Teams & Meet Nothing to install No interview recordings
How it works

One anchor, checked
every time it matters.

01

Anchor

One guided capture at the first interview becomes the candidate's only face record. It is never replaced.

02

Match

Every later round and Day One are checked against that same anchor — never against each other, so drift cannot creep in one acceptable step at a time.

03

Re-check

Brief, unannounced checks through tenure confirm the person doing the work is still the person who won the role.

Cheating has a playbook. The second camera tears it up.

  • A deepfake is engineered for the one camera it controls. Here it must survive a second one it does not — live, from the side.
  • Answers read from a screen, a coach off-camera, hands on a second keyboard: the side camera sees the desk, not just the face.
  • An earpiece can only relay what is spoken. Silent prompts appear on screen and are never said aloud.

None of this is AI detecting AI — that race the attacker wins. We change the physics, and the economics follow: defeating a camera you don't control, from an angle deepfakes aren't built for, at a moment you weren't told about, costs more than any job pays. The juice is not worth the squeeze.

512
Dimensions in the continuity template. Never an image.
1
Face record per candidate, however many rounds they sit.
0
Identity documents retained. The database rejects the write.
0.187
Highest score a different person has reached in our testing. The threshold sits at 0.60.
The problem

Hiring fraud stopped
being hypothetical.

25%
of candidate profiles worldwide are predicted to be fake by 2028.
136+
US companies infiltrated by operatives using stolen identities to win remote roles.
4
video interviews passed by one fraudulent candidate at a cybersecurity firm — plus background and reference checks.
Who this is for

Companies hiring remotely, where the person interviewed and the person who starts work are confirmed by two different processes — or by none. Background checks verify that an identity is real. They were never designed to verify that the person presenting it is the one it belongs to. That gap is where interview fraud lives, and real-time face and voice synthesis has made it cheap to exploit. Eleven governments said exactly this on 31 July 2026: their joint statement and the FBI's standing guidance warn that state operatives now use real-time deepfakes in live video interviews.

And security teams. The wrong person on Day One is not an HR record. It is an insider with credentials.

Every source behind these numbers, quoted and linked →

The premise
01 “Is this person cheating?”

A detection problem. It puts you in an arms race against generative models that improve every month and are funded better than you are. Independent testing shows detection accuracy collapsing from laboratory results to near chance on real-world material. Every win is temporary.

02 “Is this the same person?”

A continuity problem. It has a definitive answer, that answer does not degrade as the models improve, and answering it requires holding almost nothing about the person.

These are not two phrasings of one question. They are architecturally different problems — and only one of them gets easier to defend over time.
We still ruin the cheat's day. We just never have to spot them. A stand-in fails the anchor. A synthetic face has nothing to show a second camera. An earpiece cannot relay a question that was never spoken. None of this is detection. It is removing the conditions the attack needs to run.
Architecture, not policy

A promise can be broken.
A constraint cannot.

Most vendors make policy claims — we delete after 30 days, we're compliant. Those are commitments about future behaviour, resting entirely on trust in the operator. Verlinko's strongest guarantees are properties of the system: breaking one means changing the architecture, and changing the architecture is visible.

Pillar 01

Retention is refused, not discouraged

The identity record carries a database constraint that rejects any write attempting to retain a raw identity payload. Not a rule staff are asked to follow, or a setting an administrator could change under pressure. The write fails. Verlinko's own application code cannot override it.

CHECK (raw_payload_retained = false)
Pillar 02

Minimum viable identity

What persists is 512 numbers, not an image — and the system holds no name, no email and no identity document. A candidate is an opaque reference only the employer can resolve. We treat that template as regulated biometric data, because it is — encrypted, access-controlled and retention-bound, rather than described as anonymous.

Pillar 03

Re-checks leave no residue

Every continuity check after the first is compared against the original anchor and never written to storage — only the resulting score is kept. Interview twelve times and the system still holds exactly one face record: the first.

Pillar 04

Deletion is tracked, not assumed

Requesting deletion from an identity provider and confirming it happened are different things. Verlinko models purge as a state machine with retry, backoff, terminal-failure states and an attestation record — not a request fired into the dark and marked complete.

The second angle

One camera can be fed.
Two are harder.

A synthetic face injected into a video call is convincing head-on. It is also a flat plane viewed from the side. The candidate's phone, placed off-axis, is a second and independent device the attacker does not control — and it observes from an angle where a two-dimensional reproduction has far less to work with.

0°– 60°
Both cameras see near-identical views. A flat reproduction is hardest to distinguish.
60°– 110°
Depth becomes measurable, but the margin narrows under poor lighting.
110°– 120°
Guided range. Strongest depth signal while the face remains partially visible.
120°– 180°
The second camera increasingly sees only the back of a head. Identity signal is lost.
Candidates are guided into this range during setup, and the interviewer sees both feeds live. There is no automated gate here, by design: the interviewer simply asks the candidate to adjust. A human instruction, not a machine refusal. The side view also shows what the front camera never can: whether anyone else is in the room. The security page lists what is built today and what is being built next.
115°
Guided range
Flat copy clearly exposed
Drag the second camera to change the angle
Mute Question Protocol

The attack isn't detected.
It's starved.

Audio coaching — a wireless earpiece, live transcription, an assistant listening on a second device — depends on one thing: the question being spoken aloud. So the interviewer types it instead. It appears only on the candidate's screen, and the answer is spoken.

  • An earpiece cannot relay what was never spoken.
  • Transcription cannot transcribe audio that was never generated.
  • A listening assistant receives silence, and silence contains no question to answer.
  • Every prompt joins the evidence record: when it was sent, seen and answered.
Audio channelBypassed
No signal emitted
Visual channelDelivering
Response expectedSpoken aloud
Answer must originate with the candidate
Measured, not asserted

The gap between
two people and one.

A matching system is only as good as the distance between its two populations. These are measurements from Verlinko's own pipeline — not a vendor datasheet. Most vendors in this category publish no separation data at all. You can reproduce it on your own face in about a minute.

Different people — 12 pairs · 0.093 – 0.187
Same person — 31 pairs · 0.667 – 0.939
Decision threshold · 0.60
Cosine similarity between 512-dimension embeddings, faces aligned to canonical landmark positions before comparison. The unused span between 0.187 and 0.667 is the working margin.

Sample: one subject, one impostor, 43 pair comparisons. Early-stage measurement, published because it is real — not because it is complete. Multi-subject calibration is in progress and will be published the same way.
Evidence integrity

A decision you can
defend years later.

Continuity record Seal intact
check_id3f9a1c04-b7e2-4d18-9a55-0c2e
candidate_refa3f9c1…e7b2 (hashed)
similarity0.8506
threshold_used0.6000
score_bandMATCH
modelarcface/buffalo_sc
dimensions512
alignment_modealigned_5pt
compared_againstround_1_anchor
human_decisionRECRUITER_APPROVED
Every check records the rule it was judged against — not merely its outcome.
  • The threshold travels with the decision

    Rules change. Judgements shouldn't be retried under them.

    Each result stores the exact threshold in force when it was made, alongside the model and the alignment method. Tighten the standard next year and last year's decisions remain interpretable on their own terms.

  • Two gates, both real

    No session begins without a human.

    Gate one is the continuity check. Gate two is a live authorised recruiter actively approving. A leaked link on its own gets nowhere: the wrong person fails gate one, and no recruiter is standing by to pass gate two.

  • Advisory by construction

    The system never rejects anyone.

    Verlinko surfaces signals and confidence. It does not make hiring decisions and holds no mechanism to. Human review is where the decision lives — both the correct design and what regulators increasingly require of anything touching employment.

  • One anchor, always

    Every check compares to the original.

    Round four is not compared to round three. It is compared to the anchor captured at the very first session. Drift cannot accumulate one acceptable step at a time, because there are no intermediate baselines to drift through.

Continuity surface

Identity is confirmed
continuously, not once.

A check at the start of a session only proves who sat down. Verlinko re-verifies against the original anchor at every point where a substitution becomes possible.

01

Anchor capture

The one and only face record, taken at the first session and never replaced.

02

Immediately after identity verification

Confirms no substitution occurred during the document step itself.

03

At the start of every session

Including every subsequent interview round, months apart.

04

Passively during the session

Catches gradual substitution that a single opening check would miss entirely.

05

When the face leaves frame

Whoever returns is re-verified before monitoring resumes.

06

On camera or session reconnection

A disconnection is a substitution window. It is treated as one.

07

On Day One

The person who arrives is checked against the anchor from the first interview — at a stricter threshold.

08

Throughout tenure

Brief random re-checks against the original interview anchor. Minutes per quarter, scoped to roles where the wrong person at the keyboard is a security incident. The camera opens, the match runs, only the score is kept, and the moment ends. A check, never a watch.

The tenure question

You knew who they were on Day One.
What about every day since?

The day you hired this person, you signed off on who they were. Every day since, you have been trusting that nothing changed. Account handovers, outsourced shifts, and sold credentials all happen after onboarding, in the long quiet stretch no check was ever designed to cover.

Verlinko extends the interview anchor into that stretch. A tenure check is the same comparison the candidate already passed at interview, run again at a random moment, under the same consent, at a stricter threshold. Monitoring tools watch what people do. A tenure check confirms who is present, for a moment, and then stops. No image is taken or kept. A failed match asks for a retry, then routes to a human. It never accuses.

Deliberate boundaries

What Verlinko
refuses to build.

These are not gaps in a roadmap. Each was considered, argued, and rejected in writing — because building it would compromise something worth more than the feature.

Emotion or attention inference

Inferring emotional state in a workplace context has been prohibited under the EU AI Act since February 2025. Verlinko does not go near it, and will not.

One-to-many face search

Searching a face against a database is remote biometric identification, and it is what turns a verification tool into surveillance. Verlinko only ever compares one person to their own earlier self.

Session recording

Creates precisely the retention liability the product exists to eliminate. Live observation is also faster and more actionable than review after the fact.

Automated rejection

Verlinko surfaces signals and confidence. It holds no mechanism to reject a candidate, and a failed check routes to a human, never to an outcome.

Deepfake detection

An arms race with no durable winner, exited deliberately. Continuity across an entire process is a harder thing to fake than any single frame is to classify.

An email system

Knowing who to contact requires storing who they are. Employers send their own messages; Verlinko never learns a candidate's name or address.

Workforce surveillance

No session recording during tenure. No behaviour tracking. No continuous observation. A tenure check is re-attestation, not monitoring, and the difference is architectural: there is nothing running between checks.

Lockdown software

Nothing is installed, on a candidate's machine or an employee's. Verlinko is a web page that asks for the camera, runs the check, and closes. Software holding control of every employee laptop is a larger security problem than the one it was bought to answer.

Pilot programme

Anchor once.
Verify forever.

Verlinko runs alongside the video platform already in use — Zoom, Teams, Meet. No change to how interviews are conducted, no integration required to start, and nothing new for candidates to install.

We reply to every request. No newsletter, no sequence.

Patent pendingRuns beside existing video toolingAdvisory by construction