Notes from the workshop.
Reversed: GPT-4o transcription, replaced by self-hosted Whisper
We shipped the first Lina pilots on the GPT-4o transcription API and ripped it out three months later. Self-hosted Whisper large-v3 turned out faster, cheaper, and quieter on the DSGVO front.
Reversed: webhook-per-call, replaced by event sourcing
One webhook per call event collapsed the practice's n8n bridge at the 8 to 10 am peak. An append-only log fixed the retry storm in two and a half days.
Reversed: Voiceflow as backbone, replaced by a custom Node graph
Voiceflow carried the first three demos, but the block model could not hold the uncertainty gate. Replacing it with a TypeScript node graph where safeguards are middleware.
Why confidence thresholds beat intent classifiers for medical voice agents
Intent classifiers route to a handler; hand-off architectures route to a human. Why four parallel safeguards end up safer for medical calls, and the over-escalation trade-off you accept in exchange.
GDPR for voice agents: what the AVV doesn't cover
Signing the AVV is the easy 30 percent. The harder 70 percent is what the practice itself owes: patient notices, § 203 StGB authorisation, and the AI Act Art. 50 disclosure.