Nothing starts until she says yes.

Her interview starts at ten. This page follows her data through the morning: what we collect, how we lock it, who looks at it, and when it is deleted. Every line is checked against our code.

Built and tested. Not live yet.

Her camera checks send numbers, never pictures.

The checks run inside her browser. Each frame is read, then closed. What leaves her laptop is one line of numbers every 30 seconds: how many faces, how far her head turned. The interview recording is separate, kept with her consent for 180 days.

Building next. Switched off today, so it is in no check yet.

Every company gets its own lock.

If she says yes to identity matching, her face, voice and typing rhythm become numbers. No photos, no audio clips, no keys she pressed. Those numbers are locked with AES-256-GCM under a key that belongs to one company. If she applies to a second company, she enrols again, under its key.

Built and tested. Switched off until a lawyer reviews it.

Data keysExample
  • Company Akey A, version 1her templates
  • Company Bkey B, version 1nothing of hers
  • Company Ckey C, version 1nothing of hers

A root key locks each company key. They are opened in one place only: the job that checks it is still her.

Never sold, never shared with another company, never used to train models. Planned: a managed key service before our first candidate in the EU or Illinois.

Every look is written down.

Each time her template is made, matched or deleted, one line goes into a log. The line holds ids only, never her numbers. The database refuses to change or delete a line, even for us.

Access logExample
Example access log for one candidate
TimeWhatWhoResult
10:03enrol faceenrol jobok
10:21enrol voiceenrol jobok
11:03match facematch jobok
11:07match voicematch jobok
11:11match typingmatch jobok
11:15edit line 1anyonerefused

She can ask what we hold. Answering her is a line in the log too.

A flag is a question, not a verdict.

Our checks never change her score and never move her to another stage. Nothing rejects her automatically. Each check stays "In testing" until 300 honest test sessions show it flags at most 1 honest candidate in 100. We have no such numbers yet, so today every check is in testing.

How a check is tested, on the trust page →

Integrity recordExample
  • Left the page for 41 seconds at 10:27.

    In testing Does not count toward anything.
  • Pasted 412 characters at 10:44.

    In testing Does not count toward anything.

Her score, before and after these items3.0 / 4

ClearFollow-up roundReject

A person presses one of these, never a machine. She sees the same items on her data page, and can tell us her side.

Tell us your side

Then it is deleted.

Nothing of hers stays forever. Each kind of data has a date, set in one place in our code. She can delete it sooner from her data page; we answer within 30 days.

Kept until
  • Identity templates30 days after the decision
  • Backups35 days
  • Interview recording180 days
  • Transcript, feedback, resume, scores365 days
  • Consent records: what she agreed to, and when7 years

Identity templates go even sooner if she withdraws, and never last past 180 days. An ID check keeps only pass or fail, country and document type.

Ask us the hard questions.

Short answers. Where we do not know yet, we say so.

Questions
Is Crivello live?

Not yet. It is built and tested on our machines. It is not deployed, and there are no customers yet.

Do you have SOC 2?

No. We hold no security badge yet, and we will not show one we do not hold.

When will you publish false-flag rates?

Once a year, per check and per group of candidates, with the method, after a lawyer reviews them. No numbers exist yet. The thresholds stay private, so they cannot be gamed.

Do you train AI on interview data?

Identity templates: never. For everything else

Where is the data stored, and who else handles it?

Our draft list: Azure AI Foundry (AI models), Daily (video and recordings), Didit (ID check, pass or fail only), Dodo Payments (founder payments), Resend (email), Cloudflare R2 (files), Vercel (hosting), Inngest (background jobs).

Can I get a data processing agreement?

Ask us. The text needs a lawyer's review first.

Found a problem? Tell us first.

Write to content@crivello.ai with the subject "Security". Karan reads it.

Read the privacy policy. Crivello, Ludhiana, Punjab, India.