The more your team decides, the more CertAIn reads like your team.
Every time a recruiter corrects it, CertAIn absorbs that judgment and applies it to every read that follows. Your bar, your priorities, your definition of senior — in your account, on your data, never pooled with anyone else’s.
Generic AI was trained on the world’s average hiring. Not yours.
Off-the-shelf hiring AI grades every candidate against one invisible standard, an average of everyone’s idea of “qualified.” It doesn’t know you hire for trajectory over tenure, or that “5 years required” is a floor you flex. So it hands you a confident score that reflects a bar you never set. A generic score is a stranger’s opinion: useful maybe, yours never.
Four steps. You correct it once; it reads your way from then on.
- 1
You correct it.
Thumbs-down an output and a written reason is required. No silent dismissals. “This gap is fine, we hire for trajectory not tenure.” That sentence is the input: your judgment, in your words.
- 2
CertAIn synthesizes it.
It distills your team’s recent corrections into learned calibration, a plain-language read on how you judge this decision. It writes calibration; it does not train a model on your data.
- 3
It’s built to refuse the part it shouldn’t learn.
The model writing your calibration is instructed never to produce a protected characteristic or a known proxy for one — age, gender, race, national origin, religion, disability, family status, caregiving gaps, school prestige, ethnic names, salary history, criminal history. If your team’s feedback pushes on one — “prefer young grads,” “avoid candidates with gaps” — it is instructed to refuse that guidance and re-ground the calibration in job-related skills instead. A mechanical EEOC-proxy filter then checks what it wrote before anything is saved: matches are rejected, logged, and not applied, and the filter fails closed. Two controls, in that order.
- 4
Your admin approves it, and it’s applied to every future read.
Where the loop moves your numeric score adjustment, that change lands as a pending proposal — your live rankings don’t move until a tenant admin approves it in Settings → AI. No one types a number in: an admin promotes a value that already cleared the gate, or rejects it. Written calibration commits once it clears the filter. From there it’s a layer in the prompt for every AI action of that type. CertAIn now evaluates the way your team evaluates, on every candidate that follows.
The loop is batched, not instantaneous. You correct it; it incorporates the correction on the next synthesis pass, not the next millisecond. This isn’t AI that “gets smarter every second.” It’s a tool that listens to your team and remembers. That’s the honest mechanism, and it’s the one that holds up.
Calibration sharpens every place CertAIn exercises judgment.
- 01
Candidate ranking
learns what your team actually weights, so the stack-rank surfaces the people you’d surface.
- 02
Single-candidate evaluation
learns how you read a gap, a pivot, a non-linear path, so the write-up reflects your standard.
- 03
Interview prep
learns the questions your team cares about, so the prep probes what you’d probe.
- 04
AI evaluation write-ups
learns your house style for strengths and risks, so write-ups land the way your hiring managers expect.
- 05
Post-interview analysis
learns how your team reads a room, so the read after the call lines up with how you’d have called it.
It learns how you hire. It never trains a model on your data.
What CertAIn learns about your team never leaves your account. There’s no shared model that gets smarter by studying your hires, no pool your judgment flows into. The calibration lives only in your tenant, built from your corrections, used to grade exactly one company’s candidates — yours. CertAIn learns harder because it stays per-tenant: tuning to one team instead of averaging across thousands. The privacy architecture isn’t a constraint we work around. It’s the reason the learning is actually yours.
It’s built to refuse what it shouldn’t learn, then check its own work.
Most tools absorb whatever your team teaches them. Ours is built to push back. The model that writes your calibration is instructed never to produce a protected characteristic or a known proxy for one, and when your team’s feedback pushes on one — “prefer young grads,” “avoid candidates with gaps” — to refuse that guidance and re-ground the calibration in job-related skills instead. What it writes is then checked by a mechanical EEOC-proxy filter before anything is saved: matches are rejected, logged, and not applied, and the filter fails closed. Admin-authored edits run back through the same filter, so there’s no bypass.
The numeric side of calibration is a separate control: it reads only evaluation signals — no names, no résumés, no demographic fields — is bounded to ±15 points, per-recruiter, never pooled across customers, logged, and re-bounded at read time, so no stored value can move a score by more than 15 points either way. And it doesn’t move your rankings on its own: a change waits as a pending proposal until an admin approves it, and there’s no way to type a number in — an admin promotes what the gate already cleared, or rejects it.
Neither layer is infallible: a model instructed to refuse can still fail to, and a pattern-based filter can miss a phrasing it doesn’t match. They’re strong controls, not a mathematical guarantee — which is why CertAIn never auto-rejects or auto-advances a candidate: every output is a recommendation to a human decision-maker. You get personalization to your team’s bar, inside limits you can check.
The safety layer draws a line your team can’t cross.
Your team can teach CertAIn its bar. It can never teach CertAIn to cross the line the safety layer draws.
CertAIn learns to reason like your team, never to decide for it. It does not auto-reject or auto-advance anyone. Every output is a recommendation to a human, who makes the call.
Every AI read is assembled from four layers, in order of authority.
The base task
What the action is meant to do.
Locked bias guardrails
Always winsThe safety layer. Learned calibration can never override it.
Your team’s standing context
Company profile, JD overrides, the way you describe the role.
Your learned calibration
Everything the loop has absorbed from your corrections.
Layer 2 outranks Layer 4 by design.
Read it, edit it, reset it. Nothing is hidden.
Most “learning” AI is a black box: it changes, and you can’t see how. CertAIn does the opposite. In Settings → AI, admins get a Learned Calibration card for each judgment action: read it in plain English, edit a word you disagree with, reset it to baseline, or trace it back to the exact thumbs-down comments that produced it.
Weight demonstrated ownership and trajectory over years-in-seat; a non-linear path is not a penalty.
Illustrative — the real card lives in your account under Settings → AI.
The longer your team uses it, the more it reads like your team.
On day one CertAIn already reads well. It runs on your company profile and role context out of the gate. But the value builds. Every correction accumulates: the gaps you forgive, the signals you trust, the bar you actually hold. The tool stops being a smart stranger and starts reading like the person on your team who’s seen the most resumes.
The feedback button is copyable; the corrections that taught it aren’t. Your calibration is portable (read it, edit it, export the corrections behind it), but the time your team spent teaching it isn’t. The words can come with you; re-teaching them somewhere else costs that time over again. That’s not lock-in. It’s a head start that belongs to you, and grows every time your team makes a call.
Questions, answered straight.
Does this train an AI model on my data?
No. What CertAIn learns is a block of written calibration synthesized from your corrections. It lives in your tenant and is added to the prompt at read time. No shared model, no cross-customer training, ever.
Is my calibration shared with other customers?
Never. It’s built from your data, used to grade only your candidates, and never blended into a cross-customer model. Yours is yours.
Can it learn something biased?
It’s built to refuse. The model writing your calibration is instructed never to produce a protected characteristic or a known proxy for one, and is instructed to reject feedback that pushes on one — “prefer young grads,” “avoid candidates with gaps” — and re-ground the calibration in job-related skills instead. A mechanical EEOC-proxy filter then checks the result before it’s saved, rejecting and logging matches, and failing closed. On the numeric side, nothing reaches your rankings until an admin approves it. The locked safety layer always outranks learned calibration in the prompt: your team can teach CertAIn its bar, never teach it to cross the line. They’re strong controls, not a mathematical guarantee. And for aggregate patterns no single instruction reveals, there’s a check that doesn’t depend on us at all: CertAIn holds no demographic data — none, by design — so the bias-audit export gives your auditor the per-decision score and rank data to join against your own demographic records and run the impact-ratio analysis themselves.
Does it make hiring decisions on its own?
No. It does not auto-reject or auto-advance anyone. Every output is a recommendation to a human decision-maker.
What if I disagree with what it learned?
Edit it or reset it. In Settings → AI you can rewrite the calibration directly, wipe it to baseline, or open the source comments behind it. It’s your instruction; you control it.
How long until it's calibrated?
It’s useful from day one on your company profile and JD context. From there it sharpens as your team gives feedback. No training period to wait out.