From a spoken sentence to a warning in the ear.

Six stages, in a fixed order. The order is not an implementation detail — it is what makes the difference between a warning that changes the conversation and a note in a transcript nobody reads.

  1. 1

    Audio arrives from the room

    A single shared microphone streams to AssemblyAI over a WebSocket. Turn detection decides when someone has actually finished a sentence, rather than guessing from silence.

  2. 2

    Every word gets a speaker

    Streaming diarization labels words as they arrive and the dominant speaker owns the turn. Interim results update the live transcript but never reach the compliance path — nobody should be warned about half a sentence.

  3. 3

    Sensitive data is removed first

    Redaction runs before anything else touches the text. There is no code path in which raw personal data reaches the audit log, the interface, or a language model.

  4. 4

    The rule engine answers immediately

    The redacted turn is matched against the rule pack. A hit becomes a warning in the earpiece with the regulation and a sentence the professional can say out loud to correct course.

  5. 5

    A language model looks again, in parallel

    The same turn goes to a model with the conversation so far. Its findings arrive a few seconds later and are marked as such — useful, but never allowed to hold up the fast path.

  6. 6

    The agent decides whether to speak

    If the turn was addressed to it, it answers aloud from the client record in one or two sentences. Otherwise it says nothing and writes the turn to the audit log.

What happens when something fails

A live conversation cannot be paused while infrastructure recovers, so every dependency has a defined answer for being unavailable.

The language model is rate limited or downThe call continues. One nuanced finding is lost; the deterministic tier keeps catching every pattern violation with no model involved.
Speech recognition drops a connectionAudio frames are buffered and the oldest are discarded rather than letting latency grow without bound. Reconnection resumes the session.
A speaker is mis-attributedRoles are seated by speaking order and can be reassigned during the call. The audit log records the attribution that was in force at the time.
The agent is unsure it was addressedIt stays quiet. A missed cue costs one repeated question; a wrong answer in front of a client costs the professional’s credibility.

Watch the six stages run

The demo shows the pipeline end to end, with the measured latency printed beside every warning it produces.

Run a callRead the sourceChecking the backend