An AI assistant inside mental-healthcare software
Python services that turn a recorded conversation into a draft care note and appointment suggestions, with the clinician reviewing every draft.
- Role
- Software developer, Careweb, Utrecht
- Stack
- FastAPI, Whisper large-v3-turbo, WhisperX, pyannote, vLLM, React, Java / OSGi
- 1Recordinga conversation in the clinic
- 2TranscriptionWhisper, WhisperX
- 3Who said whatpyannote
- 4Draft note and appointmentsLLM served by vLLM
- 5Clinician reviewnothing is added without it
A 49-minute recording, transcribed (one local measurement)≈ 160 s
Clinicians in mental healthcare spend a lot of time writing up conversations. The assistant I work on at Careweb takes a recorded conversation, or text the clinician supplies, and drafts a care note and possible appointments from it. The clinician picks what the assistant should help with, reads the suggestions, corrects them, and chooses what goes into the care record. Nothing is added without that review.
How it fits together
The main product is an established Java application on OSGi. It keeps the documents, the permissions and the review flow. The AI runs beside it as Python services: speech becomes text, and a language model drafts the note and the appointment suggestions from that text. The result goes back to the main application for the clinician to review.
In one local measurement, a 49-minute recording was transcribed in about 160 seconds.