Vectorfy built a system that helps staff review calls for quality, compliance and customer safety. It had to work with the organisation's existing cloud provider and customer management system.
Engineering case study · Client and provider confidential
Read a call summary, see who said what, and check the transcript behind each finding. Reviewers can assess service quality and investigate privacy or safety concerns.
Detect speech and identify who spoke
Summaries, quality checks and possible compliance issues
Call insights beside the customer record
Managed releases, monitoring and GPU scheduling
A simplified view of the system. Illustrative diagram.
The required provider was slow at running the AI model. Moving to another provider was not an option.
We changed the AI framework, model, prompts and tests to meet the business needs, then deployed within the approved cloud environment.
The provider later fixed a bug that made the model run faster. Our work was adapting and delivering the system within that environment.
A leading speech model did not meet the requirements for these calls.
We fine-tuned a transcription model and built a custom audio pipeline. It detects speech, splits recordings into manageable sections and identifies who spoke when.
Identifying speakers is called diarization. It helps reviewers connect a statement or assessment to the right person.
The customer management system could not be replaced, and the integration lacked the documentation we needed.
We built the review interface inside the CRM and connected it to the call-processing services.
Staff could review transcripts and insights alongside the records they already worked with.
The custom speech models needed GPU computing, which could become expensive to run.
We used Kubernetes, the platform that runs and scales containerised services, to schedule GPU work and reduce processing costs. A separate GitOps setup, where deployment is driven from version-controlled configuration, manages releases.
The project included the application, speech processing and the infrastructure needed to run them.
A working call intelligence system in the approved cloud environment, integrated with the existing CRM. The delivery included custom speech processing and GPU scheduling to reduce running costs.
Client name and proprietary methods withheld.
