AsyncHealth

Long, multi-speaker recordings, and no early signal of who needs attention first.

Asynchronous sessions produce hours of audio with more than one voice in it. Clinicians need a clean transcript, and they need to know which sessions to open first. Neither existed as a pipeline.

Chunked, speaker-diarized transcription at 200+ sessions a week.

I built a chunked, speaker-diarized speech-to-text pipeline. It splits long recordings, labels who is speaking, and turns multi-speaker audio into structured transcripts. That automated the input layer for everything downstream.

Recall first, because a missed case costs more than an extra review.

On top of the transcripts I developed a risk-classification pipeline that analyzes sessions for symptom and suicide-risk signals. It reached 91% recall and surfaced more than ten flagged high-risk cases for clinician review each week.

Recall was the metric that mattered. A false positive costs a clinician a few minutes. A false negative is the case nobody looked at.

The model queues. The clinician decides.

Every flag lands in a review queue for a clinician. The pipeline never diagnoses, never messages a patient, and never closes a case. It changes the order in which humans look, not what they conclude.

• A chunked, speaker-diarized transcription pipeline processing 200+ patient sessions weekly into structured transcripts.
• A risk-classification pipeline for symptom and suicide-risk signals with 91% recall, surfacing 10+ flagged high-risk cases for clinician review per week.

Weekly volume

200+ sessionsTranscribed and structured automatically.

Recall

91%On symptom and suicide-risk signals.

Flagged for review

10+ / weekRouted to clinicians, never auto-actioned.