SaarHQ a Real-Time Multilingual Voice AI Sales Agent, on AWS with CloudTry

Company Overview

Wesence builds AI-powered sales tools for businesses. Its flagship product, Saar, is a multilingual voice AI agent that talks with website visitors in real time, qualifies leads and books sales meetings around the clock. Saar understands buyers who switch between English, Arabic, Hindi and Spanish mid-conversation, which helps companies capture and convert inbound interest across markets without adding sales headcount.

Customer Challenge

Many website visitors arrive late at night, from other countries, or in languages the sales team doesn’t speak, and most leave without filling in a contact form. Wesence created Saar to talk to these visitors, qualify them, and book meetings before they leave the page. Moving from prototype to a production product raised several hard problems. Voice conversations need very low end-to-end latency to feel natural, and any delay in transcribing speech, generating a reply, and speaking it back makes the conversation feel robotic. Saar’s key differentiator, following buyers who switch languages mid-sentence, requires a single multilingual speech pipeline rather than separate per-language systems. Traffic is unpredictable because it depends on customer websites across time zones, so the platform has to scale up and down quickly. As a B2B product, Saar also needs strict tenant isolation, per-customer configuration, and reliable delivery of qualified leads into each customer’s CRM. Wesence, a small team, needed to launch quickly without building and running complex GPU infrastructure in-house.

Solution Approach

CloudTry designed and built Saar’s production platform on AWS in (primary AWS Region). The embeddable voice widget is served through Amazon CloudFront and Amazon S3 and streams audio over secure WebSockets to a real-time conversation service. That service runs as containers on Amazon ECS with AWS Fargate behind an Application Load Balancer, auto-scaling with concurrent conversations.

Incoming audio is transcribed by a multilingual speech-recognition model hosted on Amazon SageMaker real-time GPU endpoints. Using one model for all languages lets Saar follow mid-sentence switches between English, Arabic, Hindi, and Spanish. Transcripts are passed to Amazon Bedrock (Anthropic Claude), which manages the dialogue, adjusts tone and formality to the visitor’s language and region, and extracts the four qualification signals: intent, budget, timeline, and sentiment. Responses are streamed back to the visitor as speech using Amazon Polly neural voices, with streaming at every stage to keep response latency low and to let visitors interrupt naturally. Bedrock Guardrails keep conversations on-topic and on-brand for each customer.

 

When a conversation ends, an event is published to Amazon EventBridge. AWS Lambda functions calculate the qualification score, book follow-up meetings through calendar integrations, and deliver lead data to each customer’s CRM through webhooks, with Amazon SQS queues and dead-letter queues ensuring retries and no lost leads. Amazon DynamoDB stores per-tenant configuration, conversation state, and lead records, partitioned by tenant for isolation. Transcripts and recordings are stored in Amazon S3 with lifecycle policies. Customers manage their agent through a dashboard secured with Amazon Cognito. Amazon CloudWatch and AWS X-Ray track latency at each stage of the voice pipeline, AWS KMS encrypts data at rest, AWS WAF protects public endpoints, and IAM least-privilege roles secure each component. Infrastructure is defined in AWS CDK and deployed through a CI/CD pipeline.

Outcomes & Impact

Wesence moved Saar from prototype to a production platform that can run live on customer websites. Visitors get natural, responsive voice conversations in English, Arabic, Hindi, and Spanish, including when they switch languages mid-sentence. The platform scales automatically with conversation volume across time zones, so Saar can engage visitors 24/7 without capacity planning. Qualified leads, scores, and booked meetings arrive in customers’ CRMs automatically and reliably. Tenant isolation and per-customer configuration let Wesence onboard new customers quickly, and Wesence’s small team can focus on the product instead of infrastructure.

Conclusion

CloudTry compared three options over three years: assembling Saar from third-party real-time voice and LLM APIs priced per minute, running the full stack on self-managed, always-on GPU servers, and an AWS architecture combining managed and serverless services. The analysis covered per-minute API fees, compute and GPU costs, storage, data transfer, implementation effort, and the engineering time needed to operate and scale each option. Third-party voice APIs were quick to prototype with, but their per-minute pricing became expensive at scale and limited control over latency and code-switching behaviour. Always-on GPU servers needed capacity for peak load and specialist operations skills. The AWS architecture uses auto-scaling SageMaker endpoints and Fargate tasks for the real-time pipeline and pay-per-use Bedrock, Polly, Lambda, and DynamoDB for everything else, so costs track conversation volume. Estimates were built with the AWS Pricing Calculator and validated against actual AWS billing after launch, and the AWS option delivered the lowest projected three-year cost per conversation.