D2 Nova is a customer communications technology company that turns high-volume call transcripts into structured, actionable business intelligence. D2 Nova needed to convert high volumes of customer call transcripts into consistent summaries, actionable follow-up tasks, and searchable customer context. ASCENDING designed and delivered a secure Generative AI architecture on AWS so business and support teams could turn every call into immediate operational insight.
BackgroundModernizing Call Intelligence for a Customer Communications Company
D2 Nova is a customer-centric communications technology company focused on improving customer interaction and operational intelligence. As conversation volume increased, teams needed a reliable way to transform unstructured transcript data into structured insight they could use in support workflows, customer engagement, and internal reporting.
The organization set a clear objective: modernize call intelligence with Generative AI while maintaining enterprise security and avoiding heavy infrastructure overhead.
The ChallengeManual Summarization and Limited Historical Call Visibility
The challenge is a widening gap between rising call volume and D2 Nova's manual, inconsistent process for turning transcripts into usable insight. D2 Nova wanted faster access to business insight from call records, but existing processes made this difficult to scale.
- Summarization and action item extraction were manual and inconsistent across teams.
- Historical conversations were hard to query at the customer level for follow-up context.
- The team needed rapid model selection and iteration without standing up complex infrastructure.
- The solution had to integrate with downstream CRM and reporting workflows while preserving data privacy controls.
AWS-Native Generative AI and SageMaker Delivery Expertise
ASCENDING is the Generative AI delivery partner D2 Nova selected to move from experimentation to secure production deployment. D2 Nova selected ASCENDING for its AWS-native delivery experience across Generative AI, SageMaker model operations, and Bedrock integration. ASCENDING combined practical implementation expertise with a production-focused architecture approach, enabling D2 Nova to move quickly from experimentation to secure deployment.
As an AWS Advanced Consulting Partner, ASCENDING brought a delivery model centered on measurable business outcomes: faster experimentation, more relevant summaries, and operationally useful call intelligence.
The SolutionBedrock and SageMaker Architecture for Call Intelligence
The solution is a multi-layered Generative AI architecture that pairs managed foundation models with domain-specific fine-tuning. ASCENDING implemented a multi-layered Generative AI solution using Amazon Bedrock for model evaluation and prompt-based workflows, plus Amazon SageMaker for fine-tuning and managed inference.
- Benchmarked foundation models in Amazon Bedrock, including Claude, Titan, and Llama, to evaluate summarization quality with secure serverless access.
- Fine-tuned a domain-specific model in Amazon SageMaker with Hugging Face trl using more than 8,000 transcript-summary pairs.
- Deployed the fine-tuned model to SageMaker Inference Endpoints for autoscaling, high availability, and secure application integration.
- Built Bedrock API workflows to extract follow-up tasks from transcripts and convert passive records into actionable operations.
- Added LangChain middleware to support dynamic customer-level queries against transcript data.
- Implemented a hybrid RAG pipeline using SageMaker-generated embeddings, customer metadata indexing, and Bedrock-based response generation.
Faster Experimentation and Actionable Call Insights
The outcome is measurable gains in speed and consistency across D2 Nova's call intelligence pipeline. D2 Nova established a scalable call intelligence capability that converted customer conversations into usable business signals faster and more consistently.
- Reduced model experimentation cycles from weeks to days by using Bedrock for rapid foundation model benchmarking.
- Improved summarization relevance with domain fine-tuning on 8,000+ real transcript-summary examples.
- Enabled immediate extraction of customer follow-up actions from call transcripts.
- Delivered customer-specific historical retrieval through metadata-tagged RAG workflows.
- Created a secure, extensible foundation for future AI-driven support and reporting use cases.
Built with AWS Generative AI and MLOps Services
Frequently Asked Questions
Which foundation models were evaluated in Amazon Bedrock for call summarization?
ASCENDING benchmarked multiple foundation models available in Amazon Bedrock, including Anthropic Claude, Amazon Titan, and Meta Llama, to compare summarization quality before selecting the best fit for domain fine-tuning in SageMaker.
How is customer call data kept secure during fine-tuning and inference?
Transcript data stays within D2 Nova's AWS environment throughout the pipeline, with the fine-tuned model deployed to SageMaker Inference Endpoints and Bedrock accessed through secure, serverless API calls rather than shared infrastructure.
Can this call intelligence pipeline integrate with existing CRM systems?
Yes. The Bedrock-based follow-up extraction workflows and LangChain middleware were designed to feed structured summaries and action items directly into downstream CRM and reporting tools without disrupting existing processes.
