Call Insights — Overview¶
Call Insights is an internal AI-powered observability and debugging system for the Atlas voice pipeline. It captures every call session in detail — audio segments, conversation turns, VAD events, STT transcripts, and latency metrics — and provides tools for diagnosing issues, comparing STT providers, and reporting transcription errors.
Components¶
| Component | Source | Purpose |
|---|---|---|
| Call Insights Service | ~/PycharmProjects/AI/call-insights-service |
FastAPI backend — session ingestion, analysis, diagnosis |
| Call Insights Frontend | ~/VSCodeProjects/Nexivo/call-insights-frontend |
React debugging dashboard |
Data Flow¶
graph LR
Atlas -->|Session report| RS[Redis Stream\ncall-insights:sessions]
RS -->|Consumer group| CIS[Call Insights\nService]
CIS -->|Store| PG[(PostgreSQL)]
CIS -->|Audio| S3[(MinIO / S3)]
CIS -->|Archive| DO[(DigitalOcean\nSpaces)]
CIS -->|Analysis| STT[STT Providers\nGroq / Deepgram / Gladia…]
CIF[Call Insights\nFrontend] -->|REST| CIS
Use Cases¶
- Browse call sessions and inspect per-turn STT transcripts, latency, and confidence
- Diagnose call issues (dropped utterances, high latency, RNNoise anomalies)
- Compare STT providers side-by-side by replaying call audio through multiple engines
- Report transcription errors with expected vs. actual output for quality tracking
- Run batch analysis across many recent calls