Context
Sales and support teams needed to review hundreds of calls a week: understand request topics, flag problematic conversations, and measure manager performance. Listening manually is impossible, and off-the-shelf services didn't offer the flexibility they needed.
Solution
Calls go to Whisper for transcription, then an LLM performs classification (categories + tags) and generates a short summary. The dashboard shows calls by manager, status, and topic; you can play back the original and view the full transcript. Labels can be edited manually — they feed back into the training set.
Stack and architecture
- Backend: Spring Boot, Spring AI (OpenAI), MongoDB.
- Speech-to-text: Whisper (via the Spring AI Transcription API).
- LLM: OpenAI (gpt-4 series) for classification and summarization.
- Frontend: Next.js, a dashboard with a list, filters, and an audio player.
- Storage: transcripts and tags in MongoDB; audio in file storage.
Role and outcome
We built the pipeline from audio ingestion to aggregates by manager and topic, designed the domain model for transcripts and tags, and implemented the dashboard UI. Analytics that previously required a team of listeners is now automatic.