Documentation Index
Fetch the complete documentation index at: https://docs.superbryn.ai/docs/llms.txt
Use this file to discover all available pages before exploring further.
Analysis pipeline
Every ingested call passes through an automated analysis pipeline that extracts quantitative metrics, evaluates transcript quality, and runs an LLM-powered audit.
Per-call metrics
Each call record tracks:
| Metric | Description |
|---|
| Duration | Total call length in milliseconds |
| Quality score | Overall quality rating |
| Sentiment | Dominant conversation sentiment |
| Avg latency | Mean response latency across turns |
| P95 latency | 95th percentile response latency |
| Cost | Total call cost in USD |
| Token usage | LLM input/output tokens, STT duration, TTS characters |
| Status | completed, failed, transferred, no_answer, busy, in_progress |
Transcript analysis
Each turn in the transcript is analyzed individually:
| Dimension | What’s measured |
|---|
| Sentiment | Per-turn sentiment and sentiment score |
| Latency breakdown | STT, LLM, and TTS latency for each turn |
| Interruptions | Whether the speaker was interrupted or interrupted the other party |
| Confidence | Speech recognition confidence score |
| Turn gap | Time between the end of one turn and the start of the next |
| Emotion | Detected emotion per turn |
Deep analysis
After processing, a comprehensive analysis record is generated:
Speaking dynamics
- Turn counts (total, agent, customer)
- Speaking time and ratios per speaker
- Words per minute per speaker
Latency
- Average latency per speaker
- Turn-by-turn latency arrays
Interruptions
- Interruption counts per speaker
- Interruption timestamps
- Agent recovery time after customer interruption
Silence
- Silence detection and metrics
- Dead air identification
AI-generated insights
- Call summary
- Chapter segmentation (title, summary, time range)
- Topic detection with frequency analysis
- Emotion analysis per segment
Call audit
An LLM-powered auditor evaluates every call across multiple dimensions.
User intent analysis
The auditor detects:
- User intent - what the caller was trying to accomplish
- Satisfaction - whether the caller’s need was met
- Reason - explanation of the satisfaction judgment
Transcript audit
Four dimensions are evaluated, each with a status and explanation:
| Dimension | Question |
|---|
| Call path | Did the call follow the expected flow? |
| Policy & guardrails | Were policies and compliance rules respected? |
| Security | Were there any security violations? |
| Tool calls | Were tool/function calls handled correctly? |
Verdict
The audit produces a verdict that classifies the call using a confusion-matrix approach:
| Verdict | Label | Meaning |
|---|
| TP | All Clear | User satisfied, agent performed well |
| FP | Blind Spot | User satisfied, but agent performed poorly |
| FN | False Alarm | User unsatisfied, but agent performed well |
| TN | Red Alert | User unsatisfied, agent performed poorly |
Observer timeline
A behaviour-level observer generates a timestamped timeline of notable events during the call. Each event has a severity level:
- Normal - expected behaviour
- Notable - worth attention but not necessarily a problem
- Deviation - something went wrong or was unexpected
Custom metric results
If you have custom metrics assigned to the agent, those evaluations also appear alongside the call analysis. Each metric shows its result (pass/fail, score, or text) with an explanation.
Export and data management
From the call log, you can:
- Export all calls matching current filters as CSV or Excel
- Batch analyze - re-queue unanalyzed calls for processing
- Search across phone numbers, agent name, and customer name
- Filter by provider, status, date range, and labels
- Sort by date, duration, cost, status, or provider
- Delete individual calls or bulk-select for deletion
