Natera Develops AI-Powered Appointment System for Cancer Patients Using Amazon Bedrock AgentCore
The developed voice assistant achieved 100% tool-calling accuracy in 500 simulations, operating at a cost of less than $0.01 per call with a perceived latency of under 7 seconds.
Natera, a diagnostics company specializing in cell-free DNA testing, has developed an automated voice appointment assistant based on Amazon Bedrock AgentCore to simplify at-home blood draw appointments for cancer patients. Migrated from Amazon ECS infrastructure to the serverless AgentCore runtime, the system replaces manual calling processes to fully automate appointment operations.
Architectural Design and Three Core Principles
Designed in collaboration between Natera and AWS, the voice assistant relies on three core architectural principles to manage real-time audio streaming. The system uses a dual WebSocket bridge pattern that separates the telephone stream from model inference. This allows infrastructure components to be modified independently without needing to redesign the entire system.
The second principle, an event-driven latency-masking technique, reduces periods of silence by providing patients with natural filler responses while data is being processed in the background. The third principle, a progressive trust model, creates a more natural dialogue flow by performing patient authentication steps gradually as the conversation progresses.
Transition from Amazon ECS to Serverless Infrastructure
Natera migrated its voice assistant workloads from Amazon ECS containers to Bedrock AgentCore runtime, a fully managed, serverless environment. During this migration process, the team decoupled the voice routing logic from the container infrastructure to make it compatible with the AgentCore interface.
To ensure long-lived WebSocket connections worked seamlessly with short-lived micro-virtual machines during the transition, a connection pooling method was implemented. Additionally, session states previously held in local container memory were moved to AgentCore memory, enhancing system reliability and reducing operational overhead.
Latency Masking and Claude Haiku Integration
The system runs a parallel filler loop so that patients do not feel like they are waiting during background tool calls. Latency periods for each operation were determined using measurements via Amazon CloudWatch logs to create automated trigger points.
During the designated wait times, short filler sentences under 15 words and appropriate to the context are generated via Amazon Bedrock using the Claude Haiku model. This method improves the user experience by ensuring the patient hears a natural assistant response instead of prolonged silences.