Nextvisit AI

What it is

Nextvisit is an ambient AI scribe for behavioral health. It listens during the visit, drafts the clinical note while the conversation is still happening, and carries that note downstream into coding, eligibility, and claim submission. I co-founded it in 2021 with Dr. Faisal Rafiq, a practicing psychiatrist.

Where it started

Faisal knew the problem from the inside. He watched colleagues lose weekends to notes, and watched good clinicians leave the field, not because they stopped caring about patients, but because the paperwork finally broke them. His own workday didn’t end when the last patient walked out. It ended at 10pm, when he closed the charting tab.

I came at it from the other side, after a few years running infrastructure inside hospitals. I knew what clinicians were relying on and where the software gave up on them.

Behavioral health makes the problem harder than most specialties, which is why we started there. A psychiatry or therapy note is narrative. There is no normal-exam default to fall back on, so the shortcuts that make primary care documentation survivable do not apply.

The parts that were actually hard

Capture nobody has to operate. Ambient means no hotword, no headset, no separate recorder app. In-person, telehealth, and hybrid visits run through the same path, so the inputs are room audio, crosstalk, and a clinician who will not stop the visit to fix your microphone. The number that mattered was whether the note coming out the far end was one a clinician would sign without editing it.

Structured notes. Clinicians want their note, in their template. SOAP, DAP, BIRP, and the documentation that addiction treatment requires are all different shapes, with different rules about what has to be present. So we treat templates as schemas: define the shape once, hand the matching JSON Schema to the model’s structured-output API, and render that same definition into the finished note. That abstraction is open source as Schematic.

Instruments as data. PHQ-9 and GAD-7 are the backbone of measurement-based care here. They have to be scored consistently and written back to the chart as structured values, or nobody can look at a panel over time and see who is slipping. We moved intake ahead of the visit with an outbound voice agent, AriaMD, in English or Spanish, so the scores existed before the clinician sat down.

EHR integration. This is the part that eats years. Every practice runs a different system with a different idea of where a note belongs. There is no single interface to build against, only a per-system negotiation, and each one has failure modes you meet in production.

Claims. A signed note is the start of getting paid. Codes suggested from the note, eligibility verified, claim scrubbed and submitted, remittance reconciled. That is X12: 837P and 837I for the claims, 270 and 271 for eligibility, 835 for remittance, and CARC/RARC codes explaining why a payer paid something other than what you billed. The standard is decades old and openly published, and every payer still reads it a little differently. We wanted same-day submission straight from the note, so that path had to be typed and testable rather than a raw HTTP client. It is open source as the Claim.MD PHP SDK.

Compliance as an engineering constraint

Nextvisit is HIPAA compliant and holds SOC 2 Type II and ISO/IEC 42001. ISO/IEC 42001 is the international AI management system standard, published 2023.

Those are architectural requirements that showed up before the first line of code. HIPAA decides which vendors you may call at all, what your logs may contain, and how long anything is kept. SOC 2 turns access control and change management into things you must evidence a year later, which means you cannot add an audit trail afterward. Those events exist from the beginning or they do not exist. ISO 42001 asks the same question about the model layer. Can you state what the system does, what data it saw, who reviewed the output, and what happens when it is wrong.

Retrofitting any of this is a rewrite.

What I would tell another engineer building here

• The model is the small part. Transcription and generation were what we understood soonest. Integration, template fidelity, permissions, audit, and billing are where the years went.
• Draft is a product boundary, not a UI state. A clinician signs the note and owns what it says. Nothing generated should land in a chart without a human putting their name on it, and the architecture should make that hard to bypass rather than merely discouraged.
• Instrument the money early. Billing has the longest feedback loop in the system. A claim defect can take weeks to come back as a denial, so anything you catch before submission is worth more than it looks.
• Find the one clinician who will tell you it is wrong. Faisal was that from the first week. Domain access is not the same as domain feedback, and the second one is what changes the roadmap.

Where it stands

The product is at nextvisit.ai. Most of my open-source time now goes to tooling for founders and small engineering teams.