Iris
the rainbow · the voice
Conversational voice AI built for healthcare.
Natural, sub-second voice conversations — with speech recognition, reasoning, and voice synthesis running entirely on servers we own, not a third-party AI cloud. Iris verifies the details that matter and remembers what you tell her.
Conversation that feels human
Iris responds fast, takes turns naturally, and carries a conversation across days — not just across one exchange.
Sub-second responses
Iris starts speaking in well under a second on typical turns — often within half a second. Replies stream sentence by sentence, so there is no dead air while she thinks.
Natural turn-taking
Iris understands when you have actually finished a thought — she waits mid-sentence and answers when you are done. Interrupt her any time and she stops and listens.
A consistent persona
Iris speaks with one warm, professional voice and character. In healthcare settings she automatically runs a stricter clinical persona that treats identifiers and dosages as sensitive by default.
Memory without limits
Iris remembers what you told her at turn three when you are at turn one hundred, and picks the thread back up when you reconnect. Conversations have no length ceiling.
Overlap-capable speech engine
A full-duplex engine that lets Iris listen and speak at the same time — backchannels, overlaps, and instant interjections. It ships only when it passes the same zero-leak safety bar as the rest of the stack.
Accurate where it counts
In regulated conversations, a mumbled digit is not a typo — so Iris verifies critical values before anything relies on them.
Verified read-back of critical details
Member IDs, dosages, dates, amounts — Iris reads them back verbatim and confirms before they are relied on. A value she cannot verify is held back and flagged, never guessed, and that read-back-before-reliance rule is adversarially tested on every change.
Dual speech-recognition cross-check
When a critical detail is on the line, two separate speech-recognition engines transcribe it in parallel and any disagreement is flagged instead of guessed. Everyday chat keeps its speed; the important fields get a second listener.
Identity, spoken
Iris knows who she is talking to — and proves it before anything sensitive is said.
Voice sign-in
Sign in by talking: your voiceprint, a spoken handle, a spoken passphrase, and a fresh liveness phrase that a pre-recorded voice cannot answer. Voice is never accepted alone — it is one factor of several, with password sign-in always available as fallback.
Speaker recognition that keeps learning
Iris recognizes enrolled voices, politely asks who new voices are, and keeps refining her ear over time — anchored to the original enrollment so learning cannot quietly wander. Sensitive read-backs are reserved for the verified account owner; other voices in the room get a value-free response instead.
Staff sign-in with two-factor
Administrators and staff sign in with email and password, protected by authenticator-app two-factor codes and single-use backup codes. Sessions are role-scoped, so people see only what their role allows.
Per-organization SSO
Each practice can bring its own identity provider: OpenID Connect single sign-on configured per organization, with just-in-time accounts that never exceed the role the organization sets.
Private by architecture
Privacy is not a policy layer on top of Iris — it is how she is built.
Self-hosted AI, end to end
Speech recognition, reasoning, and voice synthesis all run on our own private servers — the conversation itself is never handed to a third-party AI cloud for processing. The one deliberate exception: when you ask for a live web answer, a search query goes out to the web — and nothing else does. A posture designed for HIPAA-grade deployments.
A dedicated vault for personal data
Phone numbers, emails, and identifying details live in a separate hardened vault — encrypted at rest, reachable only by authenticated services, with an append-only log of every access.
Nothing kept without consent
Guests leave no transcript behind unless they explicitly opt in — and opted-in transcripts are scrubbed of identifiers before storage. Voiceprints are created only with explicit consent, and one deletion request erases a user's data across her stores, with a receipt of exactly what was removed.
More than talk
Iris is the voice of a full execution engine — she can look things up, work on real tasks, and follow through.
Live answers, mid-conversation
Ask for a fact, a score, a price, or quick math and Iris answers from the live web in a few seconds — reasoned through her engine, not read off a search page. She fills the wait naturally instead of going silent.
Hand off real work to the Zeus engine
Ask for something bigger — research, a comparison, a multi-step task — and Iris delegates it to the Zeus build engine while you keep talking. Results come back mid-call or the next time you talk.
Proactive follow-up
When a delegated job finishes while you are on the line, Iris speaks up on her own — a short update, never an interruption mid-thought. You do not have to ask twice.
Reaches you off-call
When you have stepped away, Iris can route a short pointer to text or email — never the sensitive content itself, just enough to bring you back. Delivery respects your mode: available, focus, meeting, or away.
Run it for your organization
Iris is multi-tenant from the ground up, with an operations layer built for practices — and for the developers who serve them.
Multi-tenant control plane
Every organization gets its own isolated tenant: users, roles, API keys, plans, quotas, and usage dashboards — managed from an admin console with separate operator and tenant views. Isolation is structural, not a setting. Usage is visible for operations, never billed by the token.
Streaming voice API with consent-first cloning
Developers get an industry-standard-compatible API for streaming speech synthesis and transcription, including voice cloning that requires recorded consent. Synthesized audio carries a provenance watermark.
Phone calls
Inbound and outbound telephone calling — Iris answering and placing real calls over the phone network — is scoped and in design, including the consent, disclosure, and recording safeguards regulated calling demands.
No per-minute AI rates, no character counts, no credit packs — talk to Iris all day and the price is the price. When phone calling ships, carrier charges pass straight through: set by the phone network, not by us. Final pricing is still being set — it will be under $175 a month, announced with the beta.
Iris is real, running software — not a concept video. The full conversation loop works end to end in a browser today: sub-second voice responses, verified read-back of critical details, voice sign-in, and speech, reasoning, and synthesis all running on our own hardware. We are opening a private beta soon, and we mean the label literally: the core experience runs today, telephone calling is designed but not yet live, and Iris has not yet carried production healthcare traffic. Every feature on this page is labeled live, built, or in design — because in regulated conversations, the difference between "works today" and "coming soon" is the whole point.