The human layer for regulated AI.

Every validator submits their professional license, board certification, bar admission, or academic appointment during onboarding. We independently verify each credential against the issuing authority — state medical boards, bar associations, university faculty directories. We then confirm the applicant is the license holder through Persona, with a government ID and a live selfie. Verification typically takes 1-2 business days. Unverified users cannot access any tasks. We re-verify credentials annually and when licenses are up for renewal.

The Kelva Score is a composite rating (0-100) that measures a validator's reliability across five dimensions: task accuracy (40%), inter-rater agreement (25%), reasoning quality (20%), consistency and volume (10%), and response time (5%). It's a portable credential — validators can share it on LinkedIn and professional profiles. For AI companies, the Kelva Score is the trust signal that replaces "we hope our annotators are good" with verifiable proof.

Two fundamental differences. First, transparency — most human feedback platforms are black boxes where you never see who reviewed your data. Kelva shows you exactly which credentialed professional evaluated your AI output, their accuracy history, and their reasoning. Second, the deliverable. Most providers hand back a label. Kelva hands back a signed record: who reviewed, which credential, what they decided, and when.

Each review has a clear price based on the specialty, the credential level, and the depth of review. Prices adjust for rare specialties, multi-reviewer consensus, and volume — and you always see the total before you submit. 80% goes to the reviewer, 20% to Kelva. No subscriptions, no minimums, no lock-in. Request access and we'll price it for your use case.

Kelva supports two kinds of work. Signed review: a licensed professional reviews a real output from your AI and signs off on it — approved, approved with a note, or flagged — with reasoning. Evaluation: preference ranking, hallucination detection, accuracy rating, free-form correction, multi-turn dialogue evaluation, red teaming, and full expert document review. Tasks range from 5-minute checks to 45-minute deep reviews, with pricing that scales accordingly.

It depends on the specialty and depth of review. While we onboard founding partners, we agree turnaround with each team upfront. Results come back per review as each one is signed, so you don't wait for a whole batch.

Yes. All data is encrypted in transit. Validators operate under NDA and cannot store, screenshot, or redistribute any AI outputs they review. SOC 2 Type II is on our roadmap, and we are designed to be HIPAA-compliant for medical AI customers. Enterprise tier supports dedicated validator pools where only pre-approved reviewers see their data. Full audit trails are maintained for regulatory compliance including EU AI Act requirements.

Yes — this is one of Kelva's strongest use cases. The EU AI Act requires documented human oversight for high-risk AI systems, which includes medical and legal AI. Kelva provides a full audit trail showing exactly which credentialed professional reviewed each AI output, their qualifications, their assessment, and their reasoning. This documentation is designed to satisfy the human oversight requirements that regulators will be looking for.

Because regulators are starting to ask who checked the AI. The EU AI Act requires human oversight for high-risk AI systems. In Utah, the state now lets AI renew and even write prescriptions — but only under agreements that require licensed clinicians to review the AI's decisions, every one at first and then a sample. California's SB 1120 requires a physician to make medical-necessity decisions in health-plan review, and the FDA's guidance on clinical decision support looks at whether a qualified professional can independently review the AI's output. Payers increasingly ask for the same proof. In every case, you need to show who reviewed it, with which credential, and when. That's the record Kelva produces.