Accepted — July 2026. First phone-architecture decision for the product; supersedes nothing. A future ADR will revisit conversational answering when real call data justifies it.
Resbit's core promise includes an inbound family phone number that providers, pharmacies, and vendors call instead of interrupting the caregiver. Two architectures were considered for v0.
Option A, a conversational AI agent, answers every call with a synthesized voice and triages in real time. It matches the long-term product vision shown in the marketing demo, but it is a large build with real risks at this stage. Live calls are latency-sensitive; conversations with pharmacies and medical offices are HIPAA-adjacent; and hallucination on medical content is an unacceptable failure mode for a product whose value is trust. The company is pre-revenue on a Twilio trial account.
Option B, forward-first, rings the caregiver's own phone with a whisper announcement identifying the family and caller before connecting. Answered calls are logged as handled. Unanswered calls fall to a warm voicemail. Twilio transcribes it, and the transcript lands in the call log with a digest entry. Transcripts containing an escalation keyword (defaults: discharge, hospital, emergency, urgent — caregiver-editable) trigger an SMS to the caregiver within about a minute.
Forces at play: ship speed with trusted components; zero tolerance for AI error on medical calls; the core promise still holding — no missed calls, urgent things escalate, everything logged. One more force mattered: logged real-world transcripts are exactly the training and evaluation corpus a future Option A needs.
We will build v0 forward-first. Every webhook validates the Twilio request signature (HMAC-SHA1), with missing-token treated as reject rather than bypass, and call rows are idempotent on call_sid so carrier retries cannot duplicate the log.
Positive
Negative
Neutral