What got measured, and what it cost.
A sync bug that erased the data it had just refused to read
99.3% rejectedA BLE backfill path rejected 99.3 percent of the records it asked for, then acknowledged the delete so the device wiped its own buffer. Two ordinary bugs, each survivable alone, destructive together.
Four weather APIs, one postcode, 5.7 degrees apart
5.7°C spreadSame coordinates, same minute, four providers, a 5.7 degree disagreement about the current temperature. One reported zero precipitation while the others reported rain. What that costs anything built on one upstream source.
Reversed: GPT-4o transcription, replaced by self-hosted Whisper
ripped out in 3 monthsThe first Lina pilots shipped on the GPT-4o transcription API. It came 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
fixed in 2.5 daysOne 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
3 demos, then rebuiltVoiceflow 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
4 parallel safeguardsIntent 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
the other 70%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.
Built something that broke in an interesting way? Write in, every message gets a reply.
fabi@lucidailabs.com