Inspiration
The 30 days after hospital discharge are the most dangerous part of treatment. Medication lists change and nobody checks whether the patient understood. The error that scared us into building this: a patient switched from warfarin to Eliquis who keeps taking both. Medicare already pays clinics to make these follow-up calls, and hospitals are penalized for the readmissions they prevent, yet the calls rarely happen because nurses lack the hours. The workflow is mandated, reimbursed, and unstaffed.
What it does
SafeReturn calls the patient and walks their actual discharge medication list: pickup, new meds, stopped meds, symptoms. Anything abnormal is escalated to a human by SMS within seconds, quoting the patient's own words. It never gives medical advice. It flags and escalates. That is the product.
How we built it
Twilio Programmable Voice places the call, with ConversationRelay streaming speech over a WebSocket and classic Gathete machine owns the conversation. The LLM (Groq, Llama 3.3 70B) does one job only: classify each reply as yes, no, unc question. Nothing a model generates is ever spoken to the patient. Every layer degrades gracefully: no model meansd, no Twilio means local logs. A live nurse dashboard shows the transcript and triage queue.
What we learned
Real phone tests rewrote the design. "You a negation trap for elderly patients and models alike, so the critical question became "Are you still taking it?" Speech the real bottleneck. And patients ask questions mid call, so the agent now acknowledges them and queues them for the stonewalling, while clinical questions still get a scripted deflection and an urgent escalation.
Log in or sign up for Devpost to join the conversation.