Voice Intelligence¶

Voice is the fastest intelligence input—faster than typing, faster than scanning, faster than data entry.

Archivus captures intelligence from conversations in real-time and feeds it directly into the Knowledge Graph.

The Concept¶

Traditional voice systems: Record → Transcribe → Store audio

Archivus Voice Intelligence: Speak → Transcribe in real-time → Extract intelligence → Update Knowledge Graph → Forget audio

graph LR
    SPEAK[Speak] --> STT[Real-Time STT]
    STT --> AI[Extract Intelligence]
    AI --> KG[Knowledge Graph]
    KG --> VERIFY[GOLAG Verification]
    VERIFY --> COMPLY[Compliance Log]

    STT -.Audio discarded.-> DISCARD[No Storage]

Why: Audio files are compliance liabilities. Structured intelligence is an asset.

Architecture¶

Real-Time Stack¶

Layer	Technology	Purpose
Media	LiveKit	WebRTC infrastructure, real-time streaming
STT	Deepgram	Real-time speech-to-text with diarization
TTS	Cartesia	Low-latency voice generation for responses
AI	Claude	Entity extraction, claim verification, conversation intelligence
Storage	PostgreSQL	Transcripts, extracted entities, compliance logs

Data Flow¶

User speaks
    ↓
LiveKit captures audio stream
    ↓
Deepgram transcribes in real-time (200-300ms latency)
    ↓
Claude extracts entities and claims from transcript
    ↓
Knowledge Graph updated with provenance (voice session ID)
    ↓
GOLAG agents verify extracted claims
    ↓
Compliance log records extraction + verification
    ↓
Audio stream discarded (never persisted)

Intelligence Extraction¶

During Conversation¶

What's extracted in real-time: - Entities (people, companies, dates, amounts) - Claims ("The contract is worth $1.5M") - Relationships ("John works for Acme Corp") - Action items ("Follow up with legal by Friday") - Sentiment ("Customer seems frustrated")

Example:

User: "I spoke with Sarah Mitchell from Acme Corp yesterday.
       She confirmed the Q3 revenue was up 20% to $1.5M."

Extracted:
├─ Entity: Sarah Mitchell (person)
├─ Entity: Acme Corp (organization)
├─ Entity: Q3 (temporal)
├─ Claim: "Q3 revenue increased 20%" (confidence: 0.89)
├─ Claim: "Q3 revenue was $1.5M" (confidence: 0.92)
├─ Relationship: Sarah Mitchell → employed_by → Acme Corp
└─ Source: voice_session:abc123 (provenance tracking)

Post-Session Analysis¶

After conversation ends: - Full transcript available (text only, no audio) - AI generates summary with key points - Action items extracted and assigned - Contradictions detected against existing claims - Compliance log finalized

Duration: 2-5 seconds after session ends

Compliance-First Design¶

What's Stored¶

✅ Stored (structured intelligence): - Transcript text (speaker-attributed) - Extracted entities - Verified claims - Session metadata (who, when, duration) - Intelligence extraction log - Compliance verification log

❌ Never Stored (audio): - Voice recordings - Audio streams - Biometric voice data

Why: Structured intelligence is searchable, analyzable, and has no compliance liability. Audio files are the opposite.

Regulatory Compliance¶

Regulation	Compliance Mechanism
GDPR	No biometric data stored, transcript pseudonymization
HIPAA	PII detection in transcripts, automatic redaction
SOC 2	Audit logs for all voice sessions
ISO 27001	Encryption in transit, no persistent audio storage

Audit Trail¶

Every voice session creates a compliance log:

{
  "session_id": "voice_abc123",
  "started_at": "2026-02-07T14:30:00Z",
  "ended_at": "2026-02-07T14:45:32Z",
  "duration_seconds": 932,
  "participants": [
    {"user_id": "...", "role": "host"},
    {"user_id": "...", "role": "participant"}
  ],
  "intelligence_extracted": {
    "entities": 15,
    "claims": 8,
    "relationships": 4,
    "action_items": 3
  },
  "compliance_checks": {
    "pii_detected": true,
    "pii_redacted": true,
    "sensitive_topics": ["financial_data"],
    "golag_verification": "passed"
  },
  "transcript_hash": "sha256:...",
  "hedera_anchor": "0.0.12345@1234567890.123456789"
}

7+ Year Retention: Compliance logs synced to MotherDuck/S3 Parquet for long-term storage.

Speaker Diarization¶

Concept: Automatically detect who is speaking and when.

Deepgram Diarization:

[Speaker 1, 00:00:00]: Hello, this is John from Sales.
[Speaker 2, 00:00:03]: Hi John, Sarah here from Legal.
[Speaker 1, 00:00:07]: We need to discuss the Acme contract...

Benefits: - Attribution for claims (who said what) - Conversation flow analysis - Sentiment per speaker - Action item assignment

Real-Time Features¶

Live Transcription¶

Transcript appears in UI as user speaks (200-300ms latency):

┌────────────────────────────────────┐
│  Live Transcript                    │
├────────────────────────────────────┤
│  [You, 2:31 PM]: The contract...   │
│  [You, 2:31 PM]: ...is worth $1.5M │
│                                     │
│  💡 Extracted: Contract value claim │
│  ✓ Verified by GOLAG agents         │
└────────────────────────────────────┘

Intelligence Notifications¶

Real-time notifications when intelligence is extracted:

🔍 Entity detected: "Acme Corporation"
   └─ Linked to existing entity in Knowledge Graph

📊 Claim extracted: "Q3 revenue increased 20%"
   └─ Verifying with GOLAG agents...
   └─ ✓ Verified (confidence: 0.89)

Contradiction Warnings¶

If user states something that contradicts existing knowledge:

⚠️  Potential contradiction detected

You said: "Q3 revenue was $1.2M"

But our records show:
- Source: financial_report_q3.pdf
- Claim: "Q3 revenue was $1.5M"
- Confidence: 0.95

Would you like to:
[ Investigate ]  [ Update Record ]  [ Ignore ]

Use Cases¶

1. Sales Call Intelligence¶

During customer call:
├─ Transcribe conversation
├─ Extract customer pain points
├─ Detect mentioned competitors
├─ Capture pricing discussions
├─ Identify action items
└─ Update CRM (via connector)

After call:
├─ Generate summary
├─ Create follow-up tasks
├─ Update opportunity value
└─ Flag risks to sales manager

2. Legal Consultation¶

During client meeting:
├─ Extract case facts
├─ Identify legal entities
├─ Capture dates and deadlines
├─ Detect conflicts of interest
└─ Redact PII in real-time

After meeting:
├─ Generate consultation summary
├─ Extract billable items
├─ Create deadline reminders
└─ Update case file

3. Medical Documentation¶

During patient consultation:
├─ Transcribe symptoms
├─ Extract diagnoses
├─ Capture treatment plans
├─ Identify medications
└─ Redact PHI automatically

After consultation:
├─ Generate clinical notes
├─ Update EMR (via connector)
├─ Create prescription orders
└─ Schedule follow-ups

4. Executive Meetings¶

During board meeting:
├─ Transcribe discussions
├─ Extract strategic decisions
├─ Capture action items with assignees
├─ Identify financial commitments
└─ Detect sensitive topics

After meeting:
├─ Generate meeting minutes
├─ Distribute action items
├─ Update project timelines
└─ Archive compliance log

Privacy Controls¶

Tenant Configuration¶

{
  "voice_settings": {
    "auto_redact_pii": true,
    "store_transcripts": true,
    "store_audio": false,
    "diarization_enabled": true,
    "real_time_intelligence": true,
    "compliance_mode": "hipaa" | "gdpr" | "standard"
  }
}

Before starting voice session:

┌─────────────────────────────────────┐
│  Voice Intelligence Consent          │
├─────────────────────────────────────┤
│                                      │
│  This conversation will be:          │
│  ✓ Transcribed in real-time          │
│  ✓ Analyzed for intelligence         │
│  ✓ Stored as text only               │
│  ✗ Audio will NOT be recorded        │
│                                      │
│  [ I Understand ]  [ Cancel ]        │
└─────────────────────────────────────┘

Performance¶

Metric	Specification
Transcription Latency	200-300ms
Word Error Rate	<5% (Deepgram)
Speaker Diarization	90%+ accuracy
Entity Extraction	2-3 seconds per utterance
Claim Verification	1-2 seconds (GOLAG)
End-to-End Intelligence	<5 seconds from speech to KG update

Integration with Knowledge Graph¶

Voice-extracted intelligence is first-class data:

Provenance Tracking:

Claim: "Q3 revenue increased 20%"
Source Type: voice_session
Source ID: voice_abc123
Confidence: 0.89 (lower than document, but still valid)
Extracted At: 2026-02-07T14:35:22Z
Verified By: GOLAG agents (4 votes)

Entity Linking:

Voice mention: "Sarah Mitchell"
    ↓
Search Knowledge Graph for existing entity
    ↓
Found: ent:person:sarah_mitchell:uuid
    ↓
Link voice session to entity (mention tracking)
    ↓
Update entity last_mentioned timestamp

Claim Verification:

Voice claim: "The contract is worth $1.5M"
    ↓
Check against existing claims in Knowledge Graph
    ↓
Found matching claim from contract.pdf (confidence: 0.95)
    ↓
Corroborate: Increase confidence to 0.97 (+2 sources)
    ↓
Log corroboration event

Tier Gating¶

Tier	Voice Access	Features
Free/Starter	None	No voice intelligence
Pro	None	No voice intelligence
Team	50 hours/mo	Basic transcription + entity extraction
Enterprise	Unlimited	Full intelligence + compliance + custom workflows

What This Enables¶

Traditional Voice	Archivus Voice Intelligence
Audio recordings	Structured intelligence
Manual note-taking	Automatic extraction
Post-meeting transcription	Real-time intelligence
Compliance liability	Compliance asset
Searchable by file name	Searchable by content
Isolated from other data	Integrated with Knowledge Graph

The Result¶

Fastest intelligence capture that: - Extracts knowledge in real-time - Integrates with Knowledge Graph - Verifies with GOLAG agents - Maintains compliance (no audio storage) - Enables voice-first workflows

Not "record meetings"—capture intelligence from conversations.

Intelligence at the speed of speech.