Validation infrastructure for regulated AI

“Validated” is not permanent.
So someone has to keep watching.

AI-enabled medical devices pass validation, then drift in the field — by site, by scanner, by population. VarunaForge AI is the independent layer that detects that drift and turns it into the post-market evidence regulators expect.

Independent, continuous performance monitoring for AI-enabled medical devices — across modalities and regulated markets
Performance vs. expected natural range● surveillance
DRIFT W0 W26
Point estimate Confidence interval Natural range
The problem

A model that passed validation can quietly stop performing — and nobody is watching.

Validation is a snapshot. Real-world performance moves with the population, the equipment, and the site. The regulation now treats this as the manufacturer's continuing obligation, not a one-time study.

The evidence

Measured, not theoretical.

A published real-world study of an FDA-cleared AI diagnostic, across 17 sites and 100,000+ cases, found sensitivity falling from roughly 94% to 82% — and lower still on the hardest cases. The risk is measured, not hypothetical.

Single-site blindness

One site is not the field.

One site can diverge while the rest hold. Without an independent view across sites, a manufacturer cannot tell a device problem from a site problem — opposite responses to opposite causes.

The obligation

A clearance is not a guarantee.

Across the EU and the US, regulators increasingly treat post-market performance monitoring of AI-enabled devices as a lifecycle expectation. Falling short puts market access at risk — and for a maker without the in-house infrastructure, that is existential.

What we do

The independent layer between the device and the regulator.

01 · Detect

Drift against the device's own baseline

We track sensitivity and specificity against the manufacturer's registered validation — with confidence intervals, not a single accuracy number — and flag only when performance moves outside the expected range and stays there. True drift, not natural variation.

02 · Separate

A device problem from a site problem

A view across sites is the one thing you cannot assemble from your own deployments alone. When three sites hold and one diverges, that tells you where to act — and only an independent benchmark can give you that answer.

03 · Evidence

Output your regulators trust

The result drops into your post-market evidence as a deliberately-collected, standalone component — proactive monitoring, audit-ready, and yours to file with your notified body or the FDA.

94→82%
Real-world sensitivity drop measured in a published device study
1,000+
AI/ML-enabled medical devices now authorized by the FDA
Independent
A benchmark you cannot build in-house
Audit-ready
Output aligned to notified-body and FDA expectations
Because the evidence is something you collect deliberately and continuously, it strengthens your post-market file — the kind of proactive monitoring a regulator reads as a sign of control, not a red flag.
Why proactive monitoring matters

Building an AI-enabled medical device?

The post-market performance expectation is already here, in the EU and the US alike. We can show you what independent, cross-site monitoring looks like on a device like yours — starting with your own retrospective validation data, no hospital integration required.