Intelligent call transcription and card data extraction, built on AWS

A FinTech business working with Cloud Combinator on AWS. The client is anonymised at their request.

The challenge had three focus areas.

Audio streams from the client's telephony infrastructure into Amazon Transcribe, which converts speech to text in real time with confidence scores. A completion event triggers an AWS Lambda function that stores the raw transcription in Amazon S3 and passes the text to Amazon Bedrock.

Cloud Combinator delivered a fully functional, AWS-native pipeline, deployed in eu-west-1 and defined in CloudFormation, that automatically extracts payment card details from spoken customer input. Every extraction decision is logged with a confidence score and audit trail, and the system was validated against representative the client's call data and known test card numbers across varied spoken styles.

• Real-time transcription — Amazon Transcribe converts streaming call audio into text with confidence scores, handling spoken variations of card numbers.
• AI extraction and validation — Amazon Bedrock with Claude extracts card number, expiry and CVV into validated JSON, within a thirty-second window per conversation.
• Audit-ready storage — Raw transcriptions and extracted results are stored in Amazon S3 with versioning and encryption, and tracked in DynamoDB for a complete audit trail.
• Operational visibility — CloudWatch logging, metrics and alarms flag failed or low-confidence extractions for human review rather than silent failure.
• Handover — CloudFormation templates, architecture diagrams, an API specification, operational runbooks and video walkthroughs so the client's team can own and operate the system independently.

With a production-ready pipeline and full infrastructure-as-code in its own account, the client has a foundation for AI-driven payment processing that it can extend, adding capabilities such as multi-language support and further validation rules in future phases.