cd /news/artificial-intelligence/how-stretto-built-an-ai-voice-agent-… · home › topics › artificial-intelligence › article
[ARTICLE · art-172030] src=aws.amazon.com ↗ pub= topic=artificial-intelligence verified=true sentiment=↑ positive

How Stretto built an AI voice agent with Amazon Nova 2 Sonic

Stretto, working with the AWS Generative AI Innovation Center and AWS Partner CI&T, built Tara, an AI voice agent powered by Amazon Nova 2 Sonic on Amazon Bedrock that conducts multi-phase bankruptcy intake interviews by phone and sorts leads as qualified, disqualified, or needs-more-information against attorney-defined rules. Stretto replaced its Bankruptcy.com web form — which captured only debt amount, attorney status, and contact information — with the conversational agent to collect detailed financial data such as debts, assets, vehicles, real estate, employment history, and prior filings. The team chose Nova 2 Sonic, a speech-to-speech model that processes input and generates output in a single model rather than chaining speech-to-text, language model, and text-to-speech components, to reduce latency and handle sensitive financial conversations naturally.

recommendation

For latency-sensitive voice agents, use a native speech-to-speech model rather than chaining STT, LLM, and TTS, and keep model access behind IAM and VPC endpoints for sensitive data.

more in playbook: Designing agent systems
by read10 min views3 publishedNov 18, 2026
How Stretto built an AI voice agent with Amazon Nova 2 Sonic
Image: AWS ML Blog

Artificial Intelligence #

This post is co-written with Danya Gordin from Stretto and Marcelo Wzorek from CI&T.

Before Stretto built an AI voice agent, consumers considering bankruptcy on its Bankruptcy.com site filled out a web form with a few fields: debt amount, attorney status, and contact information. These form-generated leads carried thin margins.

The gap between a web form lead and a fully qualified lead is significant. Bankruptcy attorneys place more value on a lead that includes detailed financial data: debts, assets, vehicles, real estate, employment history, and prior filings. The web form could have been extended to capture that detail, but the longer it got, the more visitors abandoned it partway through. A phone interview would have drawn the information out naturally, because people are more willing to share financially sensitive details in conversation, but staffing and training a human intake team to run those interviews was cost-prohibitive.

Stretto saw an opportunity: replace the web form with an AI voice agent that conducts the intake interview by phone. The agent would guide callers through the full bankruptcy questionnaire conversationally and extract structured data in real time. It would then sort leads against attorney-defined rules to support attorney decision-making.

The challenge was finding a voice AI solution that could handle sensitive financial conversations naturally, with low latency, clean interruption handling, and emotional awareness, at a cost that made the unit economics work.

In this post, we walk through how Stretto worked with the AWS Generative AI Innovation Center and CI&T (an AWS Partner) to build Tara, an AI voice agent powered by Amazon Nova 2 Sonic, a speech-to-speech model in Amazon Bedrock. Tara conducts multi-phase bankruptcy intake interviews and sorts leads against attorney-defined rules. We cover the architecture, the phased build from validation through production deployment, and how the team tested voice quality for sensitive conversations.

Solution overview #

Tara is an AI voice agent that conducts bankruptcy intake interviews over the phone. When a caller dials the Bankruptcy.com number, Tara picks up, introduces herself, and guides the caller through a structured conversation. She collects personal information, financial details, asset inventories, and employment history while adapting her tone to the caller’s emotional state. At the end of the call, Stretto’s system applies attorney-defined rules to sort the lead as qualified, disqualified, or needs-more-information for attorney review.

The solution is built on Amazon Nova 2 Sonic, a speech-to-speech foundation model (FM)available through Amazon Bedrock. It processes speech input and generates speech output natively in a single model rather than chaining separate speech-to-text, language model, and text-to-speech components. This architecture is designed to reduce latency and produce more natural-sounding conversations, both critical for the sensitive nature of bankruptcy intake calls. Building on Amazon Bedrock gave the team both Amazon Nova 2 Sonic and Anthropic’s Claude through one managed service, with no model infrastructure to run. It also brought AWS security and governance controls to a workload that handles sensitive financial data. Access to both models is governed by AWS Identity and Access Management (IAM). Traffic to Amazon Bedrock stays on the AWS network through virtual private cloud (VPC) endpoints (AWS PrivateLink). Amazon Bedrock doesn’t use prompts or model outputs to train models (for details, see Data protection in Amazon Bedrock). Because the models run in the same AWS environment as the voice bridge, the database, and Amazon CloudWatch monitoring, the team could secure and operate the whole solution with one set of tools. Amazon Nova 2 Sonic and Amazon Bedrock are not available in all AWS Regions. For current Region availability, see Amazon Bedrock supported Regions.

The team chose Amazon Nova 2 Sonic over the previous voice solution for three reasons:

  • Cost: Stretto reports a 95 percent reduction in per-minute voice cost compared to its previous solution, based onAmazon Bedrock on-demand pricing in us-east-1 as of April 2026.
  • Latency: In Stretto’s testing, the previous solution exhibited time-to-first-audio latency above 500 milliseconds, creating awkward s. Amazon Nova 2 Sonic is designed for low latency to support natural conversational flow.
  • Interruption handling: Callers in stressful situations need to interject, correct themselves, and ask clarifying questions. The previous solution handled barge-in poorly, often breaking the conversation flow. Amazon Nova 2 Sonic is designed to handle interruptions without breaking the session.

The following figure shows the end-to-end architecture.

A caller reaches Tara through Twilio, which opens a WebSocket connection into AWS. Traffic passes through AWS WAF and an Application Load Balancer to the voice bridge on Amazon Elastic Container Service (Amazon ECS) on AWS Fargate, which streams audio with Amazon Nova 2 Sonic (accessed through Amazon Bedrock). After the call, Amazon Simple Queue Service (Amazon SQS) with a dead-letter queue hands the transcript to an asynchronous processor. The processor uses Anthropic’s Claude 3 Haiku on Amazon Bedrock to extract structured fields, which persist in Amazon Aurora Serverless v2. Qualified leads route to attorneys by jurisdiction. Supporting services include NAT gateways, VPC endpoints (AWS PrivateLink), Amazon Elastic Container Registry (Amazon ECR), AWS Secrets Manager, and Amazon CloudWatch.

Stretto used Twilio for telephony because Tara’s intake is fully automated, with no human agents taking calls, and because Stretto wanted a direct integration with Amazon Nova 2 Sonic that kept its existing AWS backend in place. Twilio Media Streams delivers the caller’s raw audio over a WebSocket, which is the input the voice bridge needs, and CI&T had already integrated it with Amazon Nova 2 Sonic. For contact center workloads that route calls among human agents, Amazon Connect also supports Amazon Nova 2 Sonic.

The architecture uses the following AWS services:

  • Amazon Nova 2 Sonic (through Amazon Bedrock): The speech-to-speech foundation model that powers Tara’s voice conversations. It handles both speech recognition and response generation in a single pass, supporting tool calling for structured data collection during the interview.
  • Amazon Elastic Container Service (Amazon ECS) onAWS Fargate : A serverless compute engine that runs the voice bridge containers connecting the telephony layer to Amazon Nova 2 Sonic. The voice bridge tasks run continuously across three Availability Zones for high availability.
  • Amazon Aurora Serverless v2 : A PostgreSQL-compatible serverless relational database deployed in Multi-AZ configuration. Stores caller profiles, intake responses, qualification results, and session metadata. The serverless scaling of Aurora matches the variable call volume patterns.
  • Amazon Bedrock (Anthropic’s Claude 3 Haiku): Handles structured data extraction from conversation transcripts. After Amazon Nova 2 Sonic generates a transcript, Claude 3 Haiku extracts typed fields (names, dollar amounts, addresses, vehicle details) into the relational schema.
  • Amazon Simple Queue Service (Amazon SQS) with dead-letter queues: Manages asynchronous webhook processing for post-call events, preserving failed messages for retry.
  • AWS WAF , VPC Endpoints, andApplication Load Balancer : Security and networking layers that help protect API endpoints, route traffic across Availability Zones, and provide private connectivity to AWS services.

Building and validating the voice agent #

The team took a phased approach: validate voice quality first, build the production integration, then expand the questionnaire scope after proving the system worked with real callers.

Validating voice quality for sensitive conversations

Bankruptcy conversations are emotionally charged. Callers are often stressed, hesitant, or embarrassed. The voice agent needs to sound empathetic and natural, not robotic or transactional. Before writing any integration code, the team spent a dedicated week testing the voice quality of Amazon Nova 2 Sonic against several caller scenarios.

The validation covered:

  • Voice quality across caller emotional states (calm, stressed, hesitant, frustrated).
  • Time-to-first-audio latency benchmarks against the previous solution’s over 500 ms baseline.
  • Barge-in and interruption handling: can a caller interject mid-sentence without breaking the conversation?
  • Emotional tone adaptation: does the agent adjust its pace and tone when a caller sounds distressed?

This validation step produced a comparison report documenting the performance of Amazon Nova 2 Sonic against the previous solution across each dimension. The report became the go/no-go gate for proceeding to integration.

Deploying to production with real callers

With voice quality validated, the team built the end-to-end integration over three weeks. The call flow works as follows:

  1. A caller dials the Bankruptcy.com number. Twilio’s telephony layer receives the call and opens a WebSocket connection to the voice bridge running on Amazon ECS on AWS Fargate.
  2. The voice bridge streams audio bidirectionally between the caller and Amazon Nova 2 Sonic. Tara introduces herself, collects consent, and begins the intake interview.
  3. After the call, the full transcript feeds into an extraction pipeline powered by Anthropic’s Claude 3 Haiku on Amazon Bedrock. The pipeline extracts the fields from the conversation and stores them in Amazon Aurora.

The full attorney-designed intake can run 15–20 minutes, which is longer than the maximum duration of a single Amazon Nova 2 Sonic streaming session (for the current value, see Getting started with speech-to-speech). The team treated that session duration as a design parameter and implemented the session continuation pattern, an approach that carries conversation context across session boundaries. The voice bridge tracks how long each bidirectional stream with Amazon Nova 2 Sonic has been open. Shortly before the limit, it opens a new stream with the same system prompt and tool configuration and sends the conversation so far as text history. When Tara finishes her current response, the bridge switches the caller’s audio to the new stream and closes the old one, so the conversation continues smoothly for the caller.

The initial deployment covered the existing 13-field intake form, collecting the same data points the web form captured but through natural conversation. This gave the team a production baseline to measure against before expanding scope.

Expanding the intake and lead qualification

With the production pilot running, the team expanded the questionnaire to the full attorney-designed intake over four weeks. This was a significant increase in scope:

  • Over 18 intake categories with conditional branching, where the questions Tara asks depend on the caller’s previous answers.
  • Complex relational data: Vehicles (year, make, model, financing status, monthly payment per vehicle), real estate properties, individual debt line items, multiple employment and income sources, and prior bankruptcy filings.
  • Rules-based lead sorting: Stretto applies attorney-defined rules to each lead, and attorneys make the final decision. Because a voice agent collects richer, conditional data than a static web form, it gives Stretto more flexibility to apply those rules and qualify leads accurately.
  • Lead routing: Qualified leads are routed to bankruptcy practitioners, while callers who do not meet the qualification criteria are directed to alternative resources.

The Amazon Aurora schema expanded to a full relational data model supporting one-to-many relationships. A single caller record might include four vehicles, two properties, and dozens of individual debt line items.

Measuring results and optimizing

The final phase focused on collecting production data and tuning the system based on real caller behavior. The team tracked key performance indicators including call volume, average duration, intake completion rate, profile completeness, conversion rate, and drop-off points.

Insights from production data drove several optimizations: reordering questions to reduce early drop-offs, adjusting conditional branching to collect high-value fields earlier in the conversation, and modifying qualification thresholds based on attorney feedback. The following table shows two key production metrics.

| Average call duration (completed calls) | ~8 min | | Data extraction accuracy | ~96% |

Testing for quality and compliance #

Deploying a voice agent that handles sensitive financial data requires more than functional testing. The AWS Generative AI Innovation Center (GenAIIC) advisory team guided the development of a comprehensive testing framework spanning four categories:

  • Happy path testing: Complete end-to-end intake interviews with cooperative callers across the full questionnaire. These calls confirm that the over 18 categories collect accurate structured data and that the qualification rules produce correct results.
  • Adversarial testing: Callers who provide contradictory information, change answers mid-conversation, refuse to answer certain questions, or attempt to manipulate the qualification outcome. The agent needs to handle the ambiguity and keep the conversation coherent.
  • Compliance testing: Consent collection, disclosure delivery, and data handling workflows. The team verified that the agent delivers disclosures accurately and
── more in #artificial-intelligence 4 stories · sorted by recency
── more on @stretto 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/how-stretto-built-an…] indexed:0 read:10min 2026-11-18 · —