ScienceSoft’s HIPAA-compliant AI voice scheduler built on AWS

Phone scheduling in healthcare is a mess almost everyone has lived through: eight minutes on hold, then another eight to twelve minutes reciting your birthdate and insurance number to a harried receptionist. ScienceSoft, an AWS partner, is betting that voice AI can fix this without opening a new can of compliance worms, and its new system pairs Amazon's Nova Sonic speech model with Bedrock Guardrails to do exactly that.

The mechanics are straightforward once you see them laid out. Calls come in through Amazon Chime SDK, get routed to a LiveKit media server for low-latency audio, and land on ECS containers running the actual agent logic. Nova Sonic handles the back-and-forth speech-to-speech, skipping the old speech-to-text-to-LLM-to-speech relay race that makes most voice bots sound like they're on a satellite delay. Guardrails sits on top of all of it, screening every input and output in real time, redacting things like Social Security numbers, and refusing to let the model wander into medical advice.

That last part matters more than it sounds. When ScienceSoft tested the system with a patient asking for antibiotic recommendations for a sore throat, Guardrails caught it as a denied topic and the assistant redirected to booking a nurse callback instead of improvising a diagnosis. Same story with a mock prompt-injection attempt asking the bot to dump patient names — it got flagged, logged, and shut down with a canned scheduling prompt. Every one of these interventions writes to CloudWatch and CloudTrail, so a compliance officer can later see exactly what got blocked and why, with the sensitive bits redacted.

The identity check is the part that most resembles a real front-desk conversation: name, date of birth, last four of the SSN, verified against the EHR in about 20 seconds, with the bot tossing in a filler line like