Independent performance and reliability audits of dermatology and skin-imaging AI β missed cancers, unnecessary referrals, and calibration measured for every skin tone, on a published, preregistered methodology, by the team that discovered why most subgroup audits can't be trusted.
Book a 30-Minute Scoping Call See Engagement TiersOr email niya@radiantreviveaudit.com · (469) 213-8799
π Not ready to talk? Run the free, in-browser Label Reliability Check on your own dataset β
Skin-imaging AI is expanding into clinics and onto phones, but overall accuracy does not tell the whole story. A model can perform well on average and still miss important findings in individual patients. As a nurse, Niya D. Pennie understands what is at stake when a malignancy goes undetected. Radiant Revive was built to independently evaluate performance across skin characteristics and imaging conditions, helping hospitals and developers identify potential blind spots before deployment. Because confidence in AI should come from evidence that reflects the patients it will serve.
Conventional subgroup audits compute accuracy per skin tone on top of skin-tone labels assumed to be correct. Our peer-review-track research showed those labels disagree more often than they agree β which means conventional audits can pass a model that misses cancers on some skin tones, or fail one that does not. We measure the reliability of the labels first, then the performance of your model on defensible strata, with confidence intervals on everything.
Preprint: DOI 10.21203/rs.3.rs-10670899/v1
Preregistration: OSF.IO/FMSNE
Archived code: Zenodo 22039767
Audit methods v1.1: Zenodo 23149717
We never see your code, weights, or images. Tier 1 needs only your dataset's labels; Tier 2 needs only prediction outputs. Your IP never leaves your building.
Every number is computed by a release-tagged engine with SHA-256 provenance on the code, the inputs, and the report. Your audit can be re-run, byte for byte.
Founded by Niya D. Pennie, MSN, RN. Certified by the U.S. Small Business Administration as a Woman-Owned Small Business β eligible for federal set-aside procurement. Designed with reference to publicly available FDA guidance; not reviewed, qualified, or endorsed by FDA.
Federal registration: UEI EP6GX99FCN95 Β· CAGE 21GG3 Β· NAICS 541715 (primary), 541380, 541690, 541990 Β· PSC AN11, AN12, R425, R499 Β· SAM.gov active. Capability statement available on request: niya@radiantreviveaudit.com
Niya D. Pennie, MSN, RN, nurse-scientist and Principal Investigator. Preprint on skin-tone label reliability (DOI); preregistered plan (OSF); archived code and methods v1.1 (Zenodo); narrative review under peer review at the Journal of the Dermatology Nurses' Association; public sample audit published October 2026. Every audit is designed, reviewed, and signed by her.
Fixed price, fixed timeline, written scope before you sign. Start where your data is.
Each tier covers one system version, one evaluation set of up to 250,000 rows, and one operating point, delivered as a single CSV to our intake specification. Larger sets are extended at 20% per additional 250,000 rows; additional versions or sets are scoped as separate engagements at a 10% concurrent discount.
Engineered so your legal and engineering teams say yes quickly.
30 minutes. You describe the product and data; we confirm the tier and send a fixed written quote within 24 hours.
Mutual NDA + engagement letter. No data moves before both signatures.
You export one CSV to our intake spec via encrypted upload. We return a SHA-256 manifest of exactly what we received.
Version-locked engine, fixed seeds, full provenance. Validation issues are flagged back to you before analysis β never silently patched.
Findings call first, then the report: methods, tables, confidence intervals, limitations. Your data destroyed 90 days after delivery, confirmed in writing.
We have no client testimonials to show you yet, and we will not invent any. What we can show is how the engine is tested, what you receive, and what we found when we turned it on ourselves.
We ran the full audit on a public dermatology dataset and an openly available model. Read the one-page Statement of Findings (PDF) → The full 12-page report and the reproduction runner are exactly what a real engagement returns. Email niya@radiantreviveaudit.com with the subject “Sample report” Open in email app
Before any method ships, we build artificial populations where the true answer is fixed in advance and check that the engine's estimates and confidence intervals actually hold at the sample sizes real dermatology data has — especially the small darkest-skin bands. A method that fails that test does not enter the engine.
In September 2026 our own testing found that an interval we had been using was too confident on small groups, and that every interval had been treating repeat images of one patient as independent. Both were replaced, re-measured, and the changes were published with the engine. An audit firm that never finds fault in its own work is not looking.
Endpoints, thresholds, severity rules, and the language a report may use are set in a versioned public methods document (Methods v1.1, DOI 10.5281/zenodo.23149717). Nothing is chosen after we see your results. The report cites the version that governed it.
Where your tier includes it, the deliverable ships with a licensed reproduction runner β compiled and version-locked, for your internal use β that regenerates your results file from your own data and compares it to the delivered one, value by value. Your team or a regulator can check every number without trusting us.
Numbered, and verifiable by anyone
Every engagement receives a sequential registry number. If you choose to publish, a one-page Public Summary carrying the audit opinion and the SHA-256 fingerprint of the full report is posted to our audit registry, so a buyer or regulator can confirm the document in their hands is the one we issued.
Reports follow the applicable items of the TRIPOD+AI reporting guideline for external evaluations, with a conformance appendix, and open with a one-page Statement of Findings. A report is a statement of findings — never a certification, an endorsement, or a pass/fail.
If you have read the sample report and would put your name to a sentence about it, we would be glad to quote you here — with your permission, your real name, and your role.
Health systems, buyers, and regulators are starting to ask for per-skin-tone performance evidence. The vendors with an independent audit in hand will win those conversations.
Book a Scoping CallOr email niya@radiantreviveaudit.com
Radiant Revive, LLC Β· Texas woman-owned small business Β· (469) 213-8799 Β· niya@radiantreviveaudit.com
Independent research-grade auditing. Reports state methods, limitations, and confidence intervals; not a certification and not a determination of clinical safety, effectiveness, regulatory compliance, or FDA acceptability.
