AI & MLSaaSDjangoLivekitTelephonyPython

Turning a Reminder Call Prototype into a Metered Service

Delivering a reminder call is only one part of offering a reminder service. A service also needs to know whose reminder it is, when an attempt should run, what happened during delivery, and whether that outcome should affect a customer's balance.

For a project that offers scheduled reminder calls as a service, I brought those surrounding requirements into a standalone application. It has its own application data, telephony configuration, voice agent, and billing foundations, separate from an earlier medication-reminder prototype.

Deterministic delivery first

The launch-enabled mode is deterministic: prepare the reminder speech, place the outbound call, and deliver the approved message. Post-call transcription and classification help describe the outcome.

Conversational mode is recognized as future scope and remains disabled. That is an important product boundary. A service delivering a known message has a different behavioral contract from an agent improvising a conversation.

The Django application schedules and dispatches the work. A separately deployed LiveKit agent handles the outbound media workflow and returns signed outcome information.

Evidence before billing

A connected call is not automatically a successfully delivered reminder. Busy signals, no answer, incomplete playback, provider failures, and uncertain outcomes need different treatment.

The documented launch rules make complete playback to a classified human or voicemail billable, while uncertain and failed delivery remain non-billable. The purpose is to base accounting on a defined delivery outcome rather than a vague “call attempted” flag.

Retry-safe dispatch matters for the same reason. A retried job should not quietly become an unintended second charge or a duplicate call.

A private-preview foundation

The retained documentation records bounded provider tests, including an answered-human call and a corrected voicemail sequence. It also describes customer setup, local-time schedules, prepaid test checkout, and delivery/billing history.

Those are useful development milestones, but they are not a claim of broad customer operation. The recorded state has outbound gates disabled outside explicitly controlled tests.

The larger engineering lesson

Moving from a voice prototype to a service shifts attention toward boundaries: separate customer accounts, independent credentials, explicit modes, reliable outcomes, and understandable balances.

The language and speech components remain important, but the product depends just as much on what happens before and after a call. This project makes those responsibilities part of the design rather than leaving them around the edges of the demo.

Earlier project