Maya is a photorealistic digital human who greets your patients, prepares your physicians, and drafts the documentation — paired with a private medical-AI reasoner that runs on hardware inside your walls. In air-gapped mode, patient data never leaves the building.
The case for conversational AI in the clinic is overwhelming. The blocker is rarely the use case — it's the data path, and the burnout it's meant to solve.
Clinicians spend over 2 hours on documentation, record searches, and intake for every hour of face-to-face patient time. That friction is a leading driver of provider burnout and lost clinical capacity.
Most conversational and scribe vendors send patient recordings to cloud servers. For a covered entity that means third-party compliance barriers, ransomware exposure, and breaches of sensitive intake feeds.
A scribe that depends on bandwidth fails exactly when a clinic can least afford it — during an outage, a network slowdown, or an external provider's downtime.
Clinicians love the demo. Then the data-flow diagram reaches InfoSec, an arrow points out of the building, and the project quietly stalls for months.
Maya pairs a fast, warm conversational front with a slow, careful medical reasoner — both running on the clinic's own hardware. It's the dual-process design that lets her be human at the bedside and rigorous in the record.
The photorealistic talking avatar patients actually engage with — low-latency dialogue, reading cues like distress and hesitation, holding the thread of the whole visit.
A private medical-AI stack built on Google's MedGemma models — tracking symptoms against protocol, drafting notes and codes, reading images, all behind the clinic firewall.
Maya ships as a self-contained clinic appliance: the avatar host you may already own, paired with a compact on-site inference box. Plug it into the clinic LAN and it's running.
Built by a founder who spent sixteen years in data governance and regulatory compliance inside one of the world's most regulated pharmaceutical companies — so the product starts from the compliance constraint instead of bolting security on at the end.
The conversation with your security team is short, because there's no external inference path to review.
At a kiosk or exam-room tablet, Maya takes patient histories, confirms intake records, and captures the chief complaint empathetically — before the clinician walks in.
Reads local EHR records and labs into a concise huddle sheet of abnormals and gaps, delivered to the clinical workstation so the physician walks in already oriented.
Captures the consultation with a medical-tuned ASR, then drafts the SOAP note offline and populates local EHR fields — no API round-trip, no cloud transcript.
Multimodal models read X-ray, derm, fundus, CT/MRI, and pathology locally — flagging low-quality scans instead of guessing, and surfacing prior images for context.
Drafts ICD-10 and CPT codes and a prior-authorization justification from the cleaned transcript and findings — the revenue-cycle work that quietly eats clinician time.
Because inference is local, Maya keeps working when the internet drops or an external server goes down — uninterrupted, with no cloud-lag.
A real walkthrough of Maya conducting a follow-up consultation — running on local hardware, drafting as she goes.
Maya holds a natural, unhurried conversation — the kind of rapport that gets a fuller history and reduces onboarding anxiety.
While Maya speaks, the local reasoner tracks symptoms against protocol and builds the draft note in the background.
Speech, dialogue, and documentation all ran on-site. The recording never touched a third-party cloud.
Fast, empathetic dialogue up front. Deliberate reasoning behind it. A human gate before anything is finalized.
The hub reads records and labs into a concise huddle sheet of abnormals and gaps.
ReasonerMaya greets and takes history while the scribe captures it and the reasoner tracks symptoms against protocol.
Maya + scribeAny images are interpreted locally; low-quality scans are flagged rather than guessed at.
ImagingSOAP note, ICD-10/CPT codes, and prior-auth are drafted from the cleaned transcript and findings.
Back officeThe clinician reviews, edits, and signs. Only then is anything written to the EHR — and every step is audited.
ClinicianRunning inference on-premise collapses most of the HIPAA surface. With the conversation kept local, no third party touches PHI — so there's no inference-time business associate to trust in the first place.
All models run on the clinic's own hardware. In air-gapped mode, no PHI crosses the firewall — eliminating the inference-time BAA dependency entirely.
Encryption at rest and TLS in transit on the LAN; role-based access with unique IDs and automatic logoff on every node.
Every AI output is tagged with model, version, inputs, timestamp, and the clinician's accept / edit / reject decision — a defensible record of who relied on what.
Outputs are drafts. A clinician signs off before anything is finalized. These models assist clinical judgment — they don't replace it, and aren't used as clinical-grade decision-makers.
The honest fork: richer cloud voice would cross the PHI boundary and require a cloud BAA plus de-identification. So the standard deployment is air-gapped — local speech-to-text, local dialogue, local text-to-speech — keeping the boundary unbroken. Cloud voice is offered only as an opt-in tier under a signed BAA. We ship air-gapped first, by default.
"I spent sixteen years in data governance and regulatory compliance at Eli Lilly. I've seen exactly where AI projects die in a regulated environment — it's almost always the data path. We built Maya so that conversation never has to happen."
Tell us about your environment and we'll set up a working pilot — including exactly how Maya clears your security review.
The Pennar