Safa HealthFrom episodic care to continuous care
The continuous care layer for outpatient clinics. Safa answers your patients before, between and after visits, inside your clinic’s own protocol.
The conversation between visits, inside your protocol.
Wearable and CGM signal, with consent, read so your staff don’t have to.
The same control layer across chronic care.
Care comes in visits. Patients live between them
Care is continuous for eighteen minutes.
That is the whole appointment. Everything that decides whether the treatment works happens in the weeks on either side of it, where no clinic is staffed to be.
The conversation is already happening. The clinic isn’t in it
The patient
KFF Health Tracking Poll, March 2026, n=1,343
The clinic
Do et al., JAMA Network Open 2024; internal panel model
She asks at 11pm. Your protocol answers
She asks at 11pm.
Answered inside the protocol her clinic approved.
Licensed judgment goes to a human.
With the conversation, the reason and the urgency attached.
Every exchange becomes clinic data.
Versioned, timestamped, owned by the clinic.
The clinic pays. The patient never waits until morning.
Show up. Start. Stay
One layer. Three people who have to win at once.
Illustration of the staff view. Not output from a live deployment.
You keep the patient you started.
A nurse with a shorter list, not a fuller inbox. Every exchange captured as clinic-owned data.
You are never alone at 11pm.
Bounded by her own clinic’s protocol. A human the moment it matters.
Evaluation that holds up.
Blinded labeling and adjudication. An open benchmark and a published method.
Four gates. Uncertainty always resolves to a human
Refusal spine
Hard-blocks dosing, diagnosis and interactions
Triage
Emergency, urgent or routine, with confidence
Disambiguation
Bounded follow-ups, written by clinicians
Fail-safe
Any residual doubt escalates to staff
Gates are locked before unblinding: no missed emergency, and the deployed runtime must match the evaluated one.
Why here. Why now
Why here. The gap is measurable in ninety days.
Most of what makes a patient quit is tolerability, not price, and tolerability responds to contact. Slower titration with symptom support cut withdrawal almost tenfold. No clinic has a channel for that contact.
Do et al., JAMA Network Open 2024, obesity cohort n=20,217
Why now. The segment triples by 2030.
Ten million Americans take a GLP-1 today and thirty million are projected by 2030. None of the care between their visits is staffed today.
in four years, and most of that care happens between visits.
Filed, built, signed, and running
Patent pending
U.S. provisional 64/116,715, July 2026. Four claim families, 21 claims.
Safety instrument
290 triage cases, 30.7% adversarial, plus 30 multi-turn scenarios.
Design partner
Two-location Texas clinic. Design and validation gated on prerequisites.
Clinician rater study
Practicing clinicians from leading U.S. institutions rated blinded AI-agent conversations. Analysis under way.
Science accepted
Accepted at NeurIPS GenAI4Health. Multi-institution Perspective submitted to JAMIA.
Shadow mode, then paid pilots
After the release gates pass.
A public evaluation protocol, not a completed result. No claim of completed clinical validation.
She stops talking before she stops injecting
The patient most likely to quit is the one who goes quiet, and a product that only listens cannot see her leave. Next, Safa reads what she says, how she engages and, with consent, what her sensor records, and hands staff a short list of who needs a call today.
AMA International Physician Survey on Consumer Wearables, 2026
Wear patterns already predict glycemic outcomes. Whether they predict discontinuation is what our pilots measure.
Three skills. Safa needs all three
Yassen Eltayeb, founder. Author of SAFE-CARE and SAFE-CARE Bench, the framework Safa runs on.
See it inside your own protocol
Tell us which side you are on.
