WINDROSE LOGBOOK · CLINICAL INSTRUMENT
Session recording → transcript → note. No cloud.
Logbook turns a recorded therapy session into a verbatim transcript and a structured clinical note — by default entirely on one computer. Patient data never leaves the device unless you deliberately ask, and then only after pseudonymisation.
See how it works ↓Who it's for
Psychotherapists, psychologists and psychiatrists in private practice — the ones writing notes late in the evening, unwilling (and often not permitted) to send session recordings to cloud transcription, and expected to keep GDPR-grade documentation without an admin department.
How it works
Record.
Import from a recorder or record in the app, with the patient's documented consent. Audio lands in an encrypted working area.
Transcribe locally.
VERBATIMWhisper (Apple MLX) transcribes on-device, with timestamps and speaker separation.
Pseudonymise — the trust barrier.
Before any model sees the text, a local NER filter replaces people, places and organisations with tokens.
Draft the note.
AI-DRAFTA local model (Ollama) — or, if you choose, a cloud model — receives the same pseudonymised text.
Review and approve.
APPROVEDYou read the draft next to the transcript, correct it, accept it. Only then does it enter the record.
The chart.
An Obsidian vault of plain .md files; approved sessions grow a local patient profile — threads, patterns, progress.
On-device by default, cloud by choice.
a Windrose Institute principle →Hardware
Profile A — Mac, 32 GB RAM (Whisper large-v3-turbo + a 12–14B local model, sequential). Profile B — Mac, 48 GB+ (27–32B model, 32k+ context, parallel). Detected automatically at install.
FAQ
Do recordings or transcripts ever reach the cloud?
No. The default pipeline is fully local and passes an offline-only test — everything works with the network switched off.
Do I need an internet connection?
Only to download the app and models at installation. Daily work runs without a network.
What about GDPR?
You remain the data controller. Logbook records patient consent, enforces audio retention policies, and keeps an append-only audit log you can use for your records of processing. Details: /security.
What computer do I need?
A Mac with Apple Silicon and at least 32 GB of RAM for the fully local pipeline.
Can I use Claude or GPT for drafts?
Yes — as a deliberate opt-in. A cloud model only ever receives pseudonymised text; the API key lives in the macOS Keychain, and text without pseudonymisation tokens is rejected structurally before it can touch the network.
What are provenance labels?
VERBATIM marks the faithful transcript, AI-DRAFT marks a model's proposal, APPROVED marks text you accepted. Only APPROVED text enters the patient's documentation and profile.
Which session languages does Logbook support?
Today: Polish, end to end — transcription, pseudonymisation and note drafts are tuned for it. More session languages arrive only when the entire barrier, NER included, is proven for them; the changelog announces each one.