Skip to content
Pollenix Docs

Ambient AI scribe

Ambient AI scribe

Word-level audio-span citations on every scribed section. Find it inside any open encounter.

How it works

  1. getUserMedia() → MediaRecorder (16 kHz WAV) records the visit.
  2. Audio sent to the tenant’s STT provider (task=scribe) with verbose_json — returns full transcript plus word-level timestamps (start_ms, end_ms) per word.
  3. SOAP draft generated with character-span citations into the transcript — each SOAP section carries the byte range that evidences it.
  4. Click “Apply to encounter” → posts as a FHIR Composition with section.extension carrying (transcript_id, span).

Hover-to-evidence

Hover any SOAP section in the encounter detail → the supporting words highlight in the transcript pane. Click → audio playback jumps to that span. This is what makes the scribe defensible — every claim is provably derived from a specific moment in the visit.

Privacy

  • Audio is streamed to the STT provider over TLS; the raw WAV file is stored only if the tenant has scribe.retain_audio = true in settings.
  • Transcripts persist as DocumentReference resources tied to the Encounter.
  • Citations are portable FHIR primitives — any downstream consumer reading the Composition sees the same evidence chain Pollen8 does.

Continuous medication reconciliation

The scribe also pushes proposed med deltas (add / discontinue / change_dose) to the Med Reconciliation surface as the visit progresses. Provider reviews + accepts before sign-off.

Provider workflow

The scribe runs in the background. No “click to start” — opens with the encounter. Provider talks to the patient, glances at the transcript pane at sign-off, accepts the SOAP, codes flow into the encounter via the code suggester.