I built perMail for moments when a small business cannot answer: it works out who called and why, then turns the conversation into a clear owner handoff instead of leaving another transcript to process.
For a small team, answering every call is unrealistic. But a missed call can contain a new lead, an urgent customer problem, or noise. The owner needs enough context to act without replaying the whole interaction.
A useful phone product has to feel responsive while speech is still unfolding, handle imperfect turns and multilingual use, recognize urgency, and finish the operational handoff after the caller hangs up.
I built the streaming voice path, intent and urgency handling, spoken response, structured summary, business-owner notification, and the escalation path that can bring a person back into the conversation.
The default story stays at the level of the person and the work they are trying to finish.
The caller reaches a live voice flow instead of voicemail.
The system captures why they called and whether it is urgent or unwanted.
Routine calls continue; high-urgency calls can move toward a human bridge.
A structured follow-up contains the caller, reason, urgency, and next action.
The cost of being wrong changes with the call. The agent may handle routine context, but urgent or consequential situations need a clear path back to a person.
Reliability work exposed problems across audio-frame parsing, turn gating, cold starts, and stream state. Slow downstream work also competed with the part of the system the caller could feel.
Only a narrow set of operations had to happen before the next spoken response. Summaries, formatting, alerts, and owner workflow could continue after the live turn.
I kept turn handling, transcription, interpretation, and response on the conversational path, then treated structured follow-up and notification as a separate post-call path.
The product boundary became clearer: keep the caller moving now, then give the business enough evidence to take over well.
Voice loop, backend, handoff, alerts, and reliability debugging.
Urgent or consequential calls can move toward a person instead of forcing automation to continue.
The follow-up carries the caller, reason, urgency, and next action into the owner's workflow.
The technical route explains how the live call stays separate from downstream work, which failures shaped the design, and what the public evidence does—and does not—prove.
The caller-facing loop stays narrow; slower owner-facing work continues after the turn or call.
In a phone product, a stalled turn feels broken before an error message can explain it.
The agent needs bounded behavior when the audio, interpretation, or downstream handoff is uncertain.
Implementation behavior and ownership scope remain explicit without implying adoption.
Have a useful AI idea trapped in an unclear brief or brittle prototype?
Send me the rough version →