ZOHRx LLC (Gus and Rachael Valakas)

Comment submitted by ZOHRx LLC (ZOH), a Connecticut consumer health company
building general-wellness and health-education software. Our system uses
multiple AI agents to help members understand their own wearable and lab
data. We do not diagnose or treat, and our agents operate in the
general-wellness and education lane. We comment from that lane only.

1. We support the two-axis risk framework (clinical significance x level
of autonomy). It matches how responsible wellness systems are already
built: the more a system moves from providing information toward
influencing a clinical decision, the stronger the evaluation and
oversight should be. We ask FDA to keep a clear, workable boundary for
low-risk general-wellness and education functions so that proportionate
systems are not regulated as if they were diagnostic devices.

2. Disclosure should be mandatory and plain. A user should always know
they are interacting with AI, what the system can and cannot do, and
when a human or licensed clinician is the right next step. We disclose
AI use in-product and recommend FDA treat clear AI disclosure as a
baseline expectation for GenAI health functions at every risk level.

3. Outputs that touch health evidence should carry their basis. Our
agents attach an evidence grade and a confidence level to
recommendations, log how those grades perform against real outcomes
(calibration logging), and revise the recommendation when the evidence
changes. We support premarket "competency assessment" (benchmarking
plus clinical confirmation) and believe graded-evidence display and
calibration records are practical, auditable ways to demonstrate it.

4. Escalation must be designed, not improvised. Our system runs
predefined red-flag rules that stop the AI flow and direct the member
to a human or clinician partner. We recommend FDA expect documented
human-escalation pathways as part of both premarket evaluation and
postmarket monitoring for any GenAI function whose outputs could
affect a health decision.

5. Postmarket monitoring must cover the foundation model, not only the
device layer. Non-deterministic systems drift when an underlying model
changes. Developers should be expected to monitor outputs after
underlying-model updates, keep versioned records of what changed, and
report performance degradation the way they would report any other
postmarket signal.

We appreciate FDA seeking input before drafting guidance, and we are
happy to share our (de-identified) calibration and escalation design
with the agency on request.

— Gus and Rachael Valakas, Co-founders, ZOHRx LLC
2389 Main St STE 100, Glastonbury, CT 06033 · gus@zohrxllc.com