Notes from the flight deck
Releases, guides, and lessons from building open-source observability for voice AI.
Why voice AI observability has to be open source
Third-party eval tools can only see what the voice platform sees. Real observability means running inside the customer's own compliance boundary, which means open source.
Two compliance boxes: a vendor BAA vs voice AI observability you control
For healthcare voice agents, every monitoring vendor that touches call data adds another BAA and another place PHI lives. Here is how self-hosted observability changes the math.
Read"Task complete" is lying to you: how to tell if a voice agent call actually worked
Completion rates and instruction-following scores miss hangups, dead air, and failed side effects. Here is what to measure after "task complete" fails you in production.
ReadSelf-hosted observability for LiveKit voice agents: a step-by-step guide
Run SoundFlare on your own infrastructure and get traces, transcripts, cost, and AI call reviews for every LiveKit agent call, with OpenTelemetry enabled by one flag.
ReadThe booking that never was: how a silent API failure sent a patient to an empty appointment
A real production incident. Our voice agent reserved a clinic appointment, the confirmation call failed on the clinic's side, and we got the blame. Here is what we learned.
ReadSoundFlare is now free and open source
No plans, no per-call pricing, no waitlist. SoundFlare is MIT licensed and you can self-host the whole stack in one command.
ReadWhat simulated callers can't find: voice agent evaluation vs production reality
AI-generated test callers are great pre-flight checks. But most real-world voice agent failures never show up in agent-to-agent testing. Here is where each approach fits.
ReadTranscripts aren't observability: monitoring voice AI beyond the transcript
Re-reading transcripts with another LLM misses dead air, slow tool calls, and failed side effects. Here is what a real voice agent trace looks like, from STT to TTS.
Read