direction: down

operator: Operator and people {
  style.fill: "#E8F2FA"
  ui: "SvelteKit UI\nStart a meeting / Meeting"
  mic: "One shared microphone\n16 kHz mono PCM"
  people: "Multiple people\none voice at a time for enrollment"
  people -> mic: speak into
  ui -> mic: capture enrollment and meeting audio
}

app: SvelteKit server {
  style.fill: "#F0EAFB"
  auth: "Shared password\nHTTP-only session cookie"
  control: "One meeting controller\ncount, enrollment, start/stop"
  ticket: "Short-lived\naudio ticket"
  auth -> control: authenticated requests
  control -> ticket: start meeting
}

speech: Speech backend {
  style.fill: "#E7F5EE"
  ingress: "Audio WebSocket\nmeeting clock + buffer"
  profiles: "Enrollment processor\nreference embeddings"
  diar: "Nemotron-3-Diarization\nlocal NeMo or Baseten"
  verify: "Verification model\nclean live embeddings"
  match: "Identity matcher\nthreshold or unknown"
  merge: "Transcript alignment\nrevisable live state"
  ingress -> diar: same meeting PCM
  diar -> verify: select clean speaker turns
  diar -> match: anonymous speaker labels
  profiles -> match: enrolled embeddings
  verify -> match: live embeddings
  diar -> merge: timed speaker activity
  match -> merge: named speaker mapping
}

together: "Together real-time API\nParakeet transcription"

operator.ui -> app.auth: password, key, count, named samples
app.control -> speech.profiles: enrollment audio + names (private)
app.control -> speech.ingress: session config + Together key (private)
app.ticket -> operator.ui: ticket only
operator.mic -> speech.ingress: live PCM + ticket
speech.ingress -> together: same meeting PCM (server-side key)
together -> speech.merge: interim + final text
speech.merge -> operator.ui: live speaker-attributed transcript