Company Stretto deploys voice agent Tara with the Amazon Nova 2 Sonic model for bankruptcy intake conversations
The company Stretto has replaced the web form on Bankruptcy.com with a voice agent named Tara, which conducts bankruptcy intake conversations over the phone. The agent runs on the Amazon Nova 2 Sonic speech-to-speech model via Amazon Bedrock, and structured data for attorneys is extracted using the Claude 3 Haiku model.
The company Stretto, together with the AWS Generative AI Innovation Center and partner CI&T, built a voice agent named Tara, which on the Bankruptcy.com website replaces the original web form for people interested in personal bankruptcy. Instead of filling out a short form (amount of debt, whether represented by an attorney, contact details), Tara guides the caller over the phone through a multi-stage conversation and collects detailed financial data – debts, assets, vehicles, real estate, employment, and history of previous filings. According to the company, people are more willing to share sensitive financial details in a call than in a longer form, where the risk of users not completing it was growing.
The agent is built on the Amazon Nova 2 Sonic model, a speech-to-speech model available via Amazon Bedrock, which processes spoken input and generates spoken output within a single model, without separate components for speech transcription, a language model, and voice synthesis. According to the company, this setup reduces latency and makes the conversation more natural, which is important for emotionally demanding conversations about bankruptcy. The architecture combines Twilio telephony, AWS WAF and Application Load Balancer, a voice bridge on Amazon ECS/Fargate connected to Nova 2 Sonic, and after the call, Amazon SQS, which passes the recording to an asynchronous processor. This processor uses the Claude 3 Haiku model on Amazon Bedrock to extract structured data and store it in Amazon Aurora Serverless v2. Qualified leads are then sorted according to rules defined by attorneys and routed to attorneys by jurisdiction.
The team proceeded in phases: first spending a week testing the voice quality of the Nova 2 Sonic model across various caller scenarios, then building the production integration, and only afterward expanding the scope of the questionnaire. According to the company, the system has reached production deployment and is operating on Bankruptcy.com. The source further states three reasons why the team chose Nova 2 Sonic over the previous voice solution, and describes details of emotional-sensitivity testing, but these parts of the text are not available in full. See the source article for details.
Why it matters
This case shows a concrete production deployment of a speech-to-speech model for a sensitive use case, where the cost of a human intake team had previously been a barrier. For companies facing a similar problem (long forms with high user drop-off vs. costly phone-based questioning), it offers a reference architecture combining Amazon Bedrock, Twilio, and subsequent data extraction using a language model.
Two audiences, two different impacts
What this means
For individuals
Developers designing voice agents for sensitive conversations now have access to a concrete reference architecture (Twilio + Amazon Nova 2 Sonic via Bedrock + asynchronous data extraction using the Claude 3 Haiku model) that can be used as a starting pattern instead of designing from scratch.
For a business
Replacing the web form with a voice agent increased the quality and value of generated leads (detailed financial data instead of a few fields) without the need to build and train a human intake team, which the company says addresses the previously uneconomical unit costs of such operations.
ProcessesCheck the original
Event sources
only one source so far · 1 publisher, 0 independent. We count feeds from the same owner only once.