The Sentiment & Semantic Engine.
From the moment a call hits the PBX, a four-stage pipeline transforms raw audio into a living intelligence record — updated on every conversation, forever.
Four-Stage Processing Pipeline
Raw Input
Raw Audio Ingestion
Stereo PBX audio streams captured at source — no compression, no pre-processing. Every millisecond preserved for downstream analysis.
IN
PBX Stream
OUT
PCM Buffer
LATENCY
< 12ms
THROUGHPUT
50k calls/hr
data
PCM 16 kHz
Voice Agent
Voice Agent Transcription
Voice Agent converts raw audio to time-stamped token streams with speaker diarization. Word-error rate < 4% across all supported languages.
IN
PCM Buffer
OUT
JSON Tokens
LATENCY
< 340ms
THROUGHPUT
Real-time
data
JSON Tokens
Sentiment Agent
Sentiment Agent Check
Sentiment Agent scores intent, emotional valence, and churn-risk probability per utterance. Custom fine-tune trained on 12M VoIP support transcripts.
IN
JSON Tokens
OUT
Data Vector
LATENCY
< 820ms
THROUGHPUT
Parallel
data
Data Vector
PRM Engine
PRM Database Update
Data vectors are merged into the caller's longitudinal PRM profile. Running averages, trajectory slopes, and risk thresholds are recalculated on every call.
IN
Data Vector
OUT
PRM Profile
LATENCY
< 28ms
THROUGHPUT
Atomic
Raw Input
Raw Audio Ingestion
Stereo PBX audio streams captured at source — no compression, no pre-processing. Every millisecond preserved for downstream analysis.
IN
PBX Stream
OUT
PCM Buffer
LATENCY
< 12ms
THROUGHPUT
50k calls/hr
Voice Agent
Voice Agent Transcription
Voice Agent converts raw audio to time-stamped token streams with speaker diarization. Word-error rate < 4% across all supported languages.
IN
PCM Buffer
OUT
JSON Tokens
LATENCY
< 340ms
THROUGHPUT
Real-time
Sentiment Agent
Sentiment Agent Check
Sentiment Agent scores intent, emotional valence, and churn-risk probability per utterance. Custom fine-tune trained on 12M VoIP support transcripts.
IN
JSON Tokens
OUT
Data Vector
LATENCY
< 820ms
THROUGHPUT
Parallel
PRM Engine
PRM Database Update
Data vectors are merged into the caller's longitudinal PRM profile. Running averages, trajectory slopes, and risk thresholds are recalculated on every call.
IN
Data Vector
OUT
PRM Profile
LATENCY
< 28ms
THROUGHPUT
Atomic
Nov
3c
82/100
Billing enquiry resolved.
lowDec
2c
79/100
Feature request noted.
lowJan
5c
68/100
Repeated queue drops detected.
mediumFeb
8c
54/100
Tone shift: frustration markers ↑.
mediumMar
12c
38/100
Churn vocabulary detected. Alert sent.
highApr
14c
29/100
Pre-ticket window. Intervene now.
criticalPRM Alert · Pre-Ticket Window Detected
Caller #CTX-884421 trajectory predicts support ticket within 4–7 days. Proactive outreach recommended before escalation.
Proprietary Logic
People Relations Management: The Predictive Layer
Most AI reads a call and forgets it. PRM remembers every call — and uses the accumulation of sentiment data across months to predict what a customer will do before they do it.
Longitudinal Caller Mapping
Every call is linked to a persistent caller profile via phone number, account ID, or voice fingerprint. Sentiment scores accumulate into a rolling 90-day record — giving us a trajectory, not a snapshot.
Trajectory Slope Analysis
The PRM engine calculates the rate of sentiment change week-over-week. A caller whose score drops from 82 to 29 over six months has a slope that predicts ticket submission — usually 4–10 days before it happens.
Pre-Ticket Intervention Windows
When a caller enters the critical zone (score < 35, slope > −8/week), PRM triggers a proactive alert to the account manager. You call them. They never file the ticket. Churn is averted silently.
Frequent-Caller Sentiment Clustering
PRM groups callers by issue type, sentiment pattern, and contact frequency. If 12 callers from the same company all show declining scores in the same week, the system flags a systemic issue — not 12 individual ones.
Architect's Note
All models include consultative onboarding. Not sure which tier fits your network? — we'll spec the right engine for your volume and margin targets.