The engineering blueprint for capturing organisational intelligence. Five
techniques. One unified knowledge graph. Built for insurance underwriting.
Master System Architecture
graph TB
subgraph "CAPTURE LAYER"
SS["Shadow Sessions — Active, Deep"]
DA["Decision Archaeology — Active, Historical"]
SE["Scenario Elicitation — Active, Structured"]
NK["Negative Knowledge Audit — Active, Org-Wide"]
BA["Behavioral Analytics — Passive, Continuous"]
end
subgraph "PROCESSING LAYER"
TP["Transcription Pipeline"]
AP["Analysis Pipeline"]
PE["Pattern Engine"]
end
subgraph "KNOWLEDGE LAYER"
KG["Knowledge Graph — Neo4j"]
IS["Intent Specification — YAML Store"]
end
subgraph "DELIVERY LAYER"
IE["Intent Engine"]
AM["Alignment Monitor"]
end
SS --> TP
DA --> AP
SE --> TP
NK --> AP
BA --> PE
TP --> KG
AP --> KG
PE --> KG
KG --> IE
IS --> IE
IE --> AM
AM -->|"drift/freshness signals"| KG
Five capture techniques feed three processing pipelines, which populate a unified knowledge graph. The
Intent Engine enriches AI calls with expert knowledge. The Alignment Monitor detects drift and feeds
corrections back.
01
Shadow Sessions
Parameter
Value
Type
Active capture (practitioner-led)
Duration
3–5 days per expert, 2–3 experts per function
ICP Context
Sit with senior commercial underwriters during live submission evaluation in Guidewire / Duck
Creek
Captures the WHY — deep reasoning, emotional context, interpersonal judgment
Data Flow
graph LR
subgraph "INPUT"
E["Expert works on platform"]
P["Practitioner observes + asks why"]
R["Audio recording — with consent"]
end
subgraph "PROCESSING"
FW["Faster Whisper — transcription"]
SP["spaCy NLP — entity extraction"]
HE["Heuristic Extractor — custom pipeline"]
OE["Observer Events — session assembly"]
end
subgraph "OUTPUT"
HC["Heuristic Candidates"]
ST["Annotated Transcript"]
EL["Timestamped Event Log"]
end
E --> OE
R --> FW
FW --> SP
SP --> HE
P --> HE
OE --> HE
HE --> HC
FW --> ST
OE --> EL
Pipeline
Stage 1 — Audio Capture & Transcription
InputWAV/MP3 audio from practitioner's device. 2–8 hours per
session. Written consent signed.
→
ProcessingFaster Whisper (large-v3 model). Beam size 5, word
timestamps enabled, VAD filter active.
→
OutputJSON with word-level timestamps, speaker diarisation,
per-segment confidence scores.
Stage 2 — Entity Extraction
InputTranscription JSON segments
→
ProcessingspaCy (en_core_web_trf) + custom NER labels: RISK_TYPE,
THRESHOLD, DATA_SOURCE, HEURISTIC_SIGNAL, EXCEPTION_TRIGGER, RATIONALE. Sentence-transformers for
embeddings. KeyBERT for key phrases.
→
OutputAnnotated transcript with entities, key phrases,
embeddings, heuristic_type tags.
Insurance implementation: Target commercial lines UWs with 15+ years experience,
approaching retirement. Days 1–2: observe submission triage (how they decide what to quote vs. decline).
Days 3–5: deep-dive on risk score overrides, pricing adjustments, broker negotiations, exception
approvals.
02
Decision Archaeology
Parameter
Value
Type
Active capture (data-driven historical analysis)
Duration
1–2 weeks of data analysis
ICP Context
Analyse 50–100 recent underwriting decisions and their claims outcomes
Captures patterns experts can't articulate — they emerge only from aggregate data
Data Flow
graph LR
subgraph "INPUT"
PD["Policy Data — Guidewire export"]
CD["Claims Data — 3-5yr loss runs"]
UD["Underwriter Decision Logs"]
end
subgraph "PROCESSING"
DC["Data Cleaning — pandas"]
FE["Feature Engineering — domain-specific"]
PM["Pattern Mining — PrefixSpan + stats"]
IV["Invisible Variable Detection"]
end
subgraph "OUTPUT"
DP["Decision Patterns — ranked by outcome"]
IVR["Invisible Variables — report"]
HC2["Heuristic Candidates — data-derived"]
end
PD --> DC
CD --> DC
UD --> DC
DC --> FE
FE --> PM
FE --> IV
PM --> DP
IV --> IVR
DP --> HC2
IVR --> HC2
OutputFeature-enriched DataFrame (~30–50 columns). One row per
decision.
Stage 2 — Pattern Mining
Analysis
Tool
Target
Example Finding
Override pattern analysis
sklearn DecisionTreeClassifier (depth 5)
When do experts override the model?
"UWs override upward 78% when building_age > 30 AND sum_insured > £500K"
Outcome correlation
sklearn RandomForestClassifier
Which expert decisions predict fewer claims?
"broker_relationship_strength ranks #3 — not in the model"
Invisible variable detection
Feature ablation method
Find fields not in the model that predict outcomes
"time_of_day improves loss prediction 3.2% — decisions after 4pm have 18% higher loss ratios"
Expert clustering
sklearn KMeans
Group underwriters by decision style
"'Cautious' cluster has 12% lower loss ratio than 'model-trusting' cluster"
Key insight: Decision Archaeology surfaces patterns that are invisible to the experts
themselves. Nobody knows that decisions after 4pm produce worse outcomes — until you look at the data.
These become process constraints in the Intent Engine.
03
Scenario Elicitation
Parameter
Value
Type
Active capture (structured interview)
Duration
2–3 hours per expert, 3–5 experts
ICP Context
Present senior underwriters with edge-case submissions, borderline risks, novel scenarios
Output
Exception-handling heuristics, decision boundary maps, conditional logic trees
Unique Value
Captures exception handling and decision boundaries — the hardest knowledge to extract
Data Flow
graph LR
subgraph "INPUT"
SC["Scenario Library — curated edge cases"]
DA3["Decision Archaeology findings"]
end
subgraph "CAPTURE"
VI["Interview Recording"]
SR["Structured Response Form"]
end
subgraph "PROCESSING"
FW2["Faster Whisper"]
DB["Decision Boundary Mapper"]
CL["Conditional Logic Extractor — LLM"]
end
subgraph "OUTPUT"
EH["Exception Heuristics"]
DBM["Decision Boundary Maps"]
CLT["Conditional Logic Trees"]
end
SC --> VI
DA3 --> SC
VI --> FW2
SR --> DB
FW2 --> CL
CL --> EH
DB --> DBM
CL --> CLT
Scenario Library (Insurance UW)
Category
Purpose
Example
Borderline Risk
Submissions on the approve/decline boundary
Restaurant chain: clean history but one flood-zone location, no fire suppression upgrade, sum
insured £2.1M, model score 68 (borderline)
Novel Risk
Risks the expert hasn't seen before
Indoor vertical farming facility, £5M, no comparable loss data, converted 1970s warehouse
Exception Handling
When standard process doesn't apply
15-year client, £3M annual premium, major claim from contractor negligence, model recommends 35%
increase
What-If Variations
Map decision boundaries by varying one factor
"Same submission, but building is 50 years old." Record exactly when the decision flips.
Decision Boundary Mapping
For each what-if series: vary ONE parameter, record expert's decision at each threshold, plot the
inflection point where the decision changes, and capture the stated reason for the boundary.
// Example: Scenario Elicitation heuristic
{
"heuristic_id": "HEU-SE-001",
"type": "exception_override",
"description": "Long-standing clients (>10yr, >£1M
premium) with single claim from third-party negligence: cap increase at 15%",
"decision_boundary": {
"tenure_threshold": "10 years — below this, follow
model",
"claim_count_threshold": "2+ claims in 5yr — follow model
regardless",
"premium_threshold": "< £500K annual — follow model, lower
relationship value"
},
"cross_expert_agreement": "2 of 3 experts agreed",
"confidence": 0.72
}
04
Negative Knowledge Audit
Parameter
Value
Type
Active capture (org-wide survey + interviews)
Duration
1–2 weeks
ICP Context
Uncover failed AI initiatives, broken processes, known workarounds in underwriting
Prevents AI from repeating known mistakes. Captures knowledge that is actively AVOIDED.
Data Flow
graph LR
subgraph "INPUT"
SV["Anonymous Survey — all UW staff"]
FI["Follow-Up Interviews — 30 min each"]
PM2["Post-Mortem Reports"]
end
subgraph "PROCESSING"
SA["Survey Analysis — pandas"]
TA["Thematic Analysis — LLM-assisted"]
WD["Workaround Detection"]
end
subgraph "OUTPUT"
AP["Anti-Pattern Library"]
WR["Workaround Registry"]
IM["Institutional Memory Map"]
end
SV --> SA
FI --> TA
PM2 --> TA
SA --> AP
TA --> AP
WD --> WR
TA --> IM
Survey Domains
Domain
Questions
Failed Initiatives
What AI/automation has been tried and didn't work? What guidelines look good on paper but fail
in practice?
Workarounds
What process steps do you regularly skip? Do you use personal spreadsheets alongside the
official system?
Knowledge Gaps
What questions from juniors are hardest to answer? What does the org understand least?
Assumptions
What "truths" are outdated? What risks is the org systematically mispricing?
// Example: Anti-pattern from failed automation
{
"anti_pattern_id": "AP-001",
"type": "failed_automation",
"description": "Auto-triage model (hard score cutoff at 50)
abandoned after 3 months — was declining profitable risks with unusual profiles",
"lesson": "Never use model scores alone for decline
decisions. Expert review required for scores 30–70.",
"ai_mitigation": "Intent Engine constraint: no automated
decline without expert review for scores 30–70"
}
05
Expert Behavior Analytics
Passive capture — automated, continuous, zero-intrusion. The UEBA-inspired core of the Tacit engine.
Parameter
Value
Type
Passive capture (automated, continuous)
Duration
Continuous after deployment
ICP Context
Observer Agent on Guidewire/Duck Creek captures how underwriters actually assess risk
The "Why?" button: When an expert makes an unusual decision, the extension prompts:
"That's an interesting choice — want to note why?" Annotations become the highest-value nodes in the
knowledge graph. They bridge passive capture with active elicitation.
Pattern Engine Detail
Stage
Tool
Function
Session Assembly
Custom Python (asyncio)
Group events by expert + time proximity (gap > 5 min = new session). Calculate: duration,
screen sequence, time per screen, overrides, annotations.
Baseline Modeling
River (CluStream + ADWIN)
Incremental learning per expert. Features: session duration, screen count, override count, data
source set. Alerts on behavioral change points.
graph TD
E["Expert"] -->|contributes| H["Heuristic"]
H -->|applies_to| D["Decision Domain"]
H -->|has_condition| C["Condition Set"]
H -->|produces| A["Action"]
H -->|conflicts_with| H2["Other Heuristic"]
D -->|regulated_by| R["Regulation"]
I["Intent Spec"] -->|constrains| D
The Observer SDK is a zero-dependency TypeScript library that runs silently in the expert's browser,
capturing the implicit behavioural signals that reveal tacit knowledge — without
requiring the expert to do anything differently. It records not what they type, but how they
think.
Parameter
Value
Type
Passive capture (continuous, zero-effort)
Footprint
<20KB gzipped, <1% CPU
Privacy
SHA-256 hashing + PII redaction before data leaves the browser
Captures what experts cannot articulate — the patterns embedded in their actions
Architecture
graph TB
subgraph "Browser — Expert's Application"
DW["DOM Watcher"]
CA["Cursor Analyzer"]
HA["Hesitation Analyzer"]
JT["Judgment Tracker"]
JM["Journey Mapper"]
end
subgraph "Processing"
PE["Privacy Engine"]
SM["Session Manager"]
BB["Baseline Builder"]
AD["Anomaly Detector"]
end
subgraph "Output"
EB["Event Buffer"]
OQ["Offline Queue — IndexedDB"]
TP["Transport — WebSocket"]
end
DW --> PE
CA --> PE
HA --> PE
JT --> SM
JM --> SM
PE --> SM
SM --> BB
BB --> AD
SM --> EB
AD --> EB
EB --> OQ
EB --> TP
Cursor & Mouse Behavioural Analytics
The SDK includes a dedicated Cursor Analyzer with 7 real-time detectors, grounded
in Meidenbauer et al. (PMC10084322) research on mouse movements reflecting personality traits and
cognitive states:
Detector
Signal
What It Reveals
Hover Linger
Cursor stationary >1.5s over element without click
Uncertainty, curiosity, or searching for information not immediately visible
Dead Zones
50×50px grid cells with zero mouse activity
Page areas that are invisible or unimportant to the expert — potential UI clutter
Slow Deliberate
Speed <50px/s sustained >1s
High cognitive load — expert is reading intently, analysing, or struggling with complexity
Erratic Movement
Speed >800px/s with rapid direction changes (>90°)
Frustration, impatience, or disorientation — the expert may be lost in the interface
Idle Cursor
<5px/s for >2s
Reading, thinking, or watching content — deep internal processing
Two-Step Targeting
Fast move (>600px/s) → slow corrective (<100px/s)
Goal-oriented decisive action — the expert knows exactly what button they want
Text Selection
Highlighting text on screen
Cognitive anchoring on complex or non-structured content to maintain reading focus
Fixation
Micro-movements within 25px for >250ms
Deep attention and focused reading (per Meidenbauer et al. methodology)
Hesitation & Deliberation Patterns
Pattern
Detection
Significance
Pause Before Commit
>10s inactivity before a field change or submission
The expert is weighing a difficult decision — high deliberation moment
Undo/Redo
A → B → A value pattern in a field
Second-guessing, testing alternatives, or reconsidering initial judgment
Extended Idle
>30s of no DOM events
Expert has stepped away, is thinking deeply, or has switched context
Rapid Scan
3+ field focuses in <2s intervals
Quick triage — expert is scanning for a specific signal rather than reading sequentially
Expert for the User
The SDK is designed so that the expert — typically a non-technical domain professional — has
an effortless experience:
Step 1Log in as usual. Nothing to install.
→
Step 2Accept the one-time consent banner.
→
Step 3Work normally. SDK runs invisibly in the background.
→
OptionalClick the floating "Why?" button to explain a tricky
decision.
Privacy guarantee: Expert identity is immediately SHA-256 hashed. Sensitive fields
(SSN, DOB, etc.) are redacted to [REDACTED]before data leaves the browser.
No screenshots, no keylogging, no tracking outside the work application.
10
Socially Interactive Agent (SIA)
The SIA is a digitally embodied colleague — warm, competent, and contextually aware —
that conducts structured reflection dialogues with domain experts. It captures what people
know but cannot articulate by engaging them in natural, trust-building conversations
that surface deep experiential knowledge.
Parameter
Value
Type
Active capture (AI-guided conversational reflection)
Duration
15–25 minutes per session, yielding 2–4 codified tacit rules
Trigger
Scheduled, anomaly-detected, or expert-requested
Output
Codified heuristics, decision boundary maps, training scenarios
Unique Value
Captures the WHY behind expert behaviour — the reasoning that the Observer SDK alone cannot
surface
Dual-Channel Knowledge Capture
The SIA and Observer SDK form a complementary system:
Certainty Index: The SIA fuses all four channels in real-time to compute a
continuous score (0–1) estimating how confident the expert is in what they're currently saying.
When the index drops, the SIA knows to probe deeper.
Chain-of-Thought Dialog Strategy
The SIA does not passively respond to prompts. It drives the conversation through
five phases designed to excavate tacit knowledge layer by layer:
Phase
Purpose
Example
1. Rapport & Anchoring
Establish trust via warmth and recognition of expertise
"I noticed you've been handling coastal property risks for about 8 years — that's deep
experience. How do you think about flood proximity differently from what the models show?"
2. Situated Recall
Trigger episodic memory with specific case probing
"Can you walk me through a specific case where you overrode the model's flood risk score? What
did you see that the model missed?"
3. Cognitive Excavation
Decompose intuition into teachable steps
"You mentioned the satellite view. If a junior underwriter were sitting next to you, what would
you point at on the screen?"
4. Contrastive Elicitation
Map decision boundaries via counterfactuals
"What if the property were 200 meters further east, past the ridge? Would your assessment
change?"
5. Validation & Codification
Mirror the extracted heuristic back for confirmation
"Your rule seems to be: 'For coastal properties, check satellite for elevation features like
ridges not captured in flood zone maps.' Does that sound right?"
Trust & Persona Engine
Trust is not a feature — it is the prerequisite for tacit knowledge sharing.
The SIA dynamically calibrates its persona along two research-backed dimensions:
Dimension
How the SIA Expresses It
Adaptation Over Time
Warmth
Active listening cues (nodding, "I see"), empathetic phrasing, remembering previous
conversations
High in early sessions → balanced as trust establishes
Competence
Correct domain jargon, referencing regulations, informed counterfactuals, citing expert's own
past decisions
Moderate initially → increases as SIA proves itself knowledgeable
RAG-Powered Organisational Memory
The RAG engine gives the SIA institutional context via four real-time data sources:
Document CorpusPolicies, guidelines, regulatory texts, training
manuals
Knowledge GraphCodified tacit rules from previous SIA sessions
Behavioural BaselinesPer-expert fingerprints from Observer SDK
Case HistoryPast decisions, outcomes, and expert annotations
SIA Embodiment
Research shows people disclose 40% more information to embodied agents vs text interfaces
(Krämer et al., 2018). The SIA uses:
Modality
Implementation
Purpose
Visual
Realistic avatar with real-time lip sync, micro-expressions, gaze direction
Social presence triggers natural storytelling and "thinking aloud"
Vocal
Neural TTS with prosodic variation, paced pauses, matched cadence
Strategic pauses before important questions signal that the question matters
Knowledge Externalisation Output
// Example: Codified tacit rule from a SIA session
{
"rule_id": "TK-2026-0847",
"source_expert": "af901a8...",
"confidence": 0.92,
"domain": "coastal_property_underwriting",
"rule": "When model flags high flood risk for coastal
properties,
check satellite imagery for geological ridges. Properties above
the flood plain on ridges are systematically over-scored.",
"boundary_conditions": [
"Only applies within 2km of coastline",
"Does not apply if in designated SFHA zone",
"Ridge must be visible in satellite/topographic data"
],
"validation": {
"expert_confirmed": true,
"corroborated_by": 2,
"outcome_validated": true,
"historical_accuracy": 0.87
}
}
Technology Stack
Layer
Technology
Rationale
Foundation LLM
GPT-4o / Claude 3.5 / Gemini 2.0
Conversational intelligence with long-context for session continuity
CoT Engine
Custom prompt chains + state machine
Structured dialog strategy adapted to expert responses
Relationships between experts, rules, cases, and domains
ASR
Whisper v3 / Deepgram Nova-2
Real-time speech recognition with domain vocabulary
TTS
ElevenLabs / Azure Neural TTS
Expressive speech with prosodic control
Avatar
Ready Player Me + NVIDIA Audio2Face
Real-time lip sync and expression mapping
Emotion Detection
MediaPipe + custom classifier
Micro-expression and prosody analysis for Certainty Index
The feedback loop: Observer SDK detects an anomaly → SIA is triggered to probe
the expert's reasoning → New rule is codified and added to the Knowledge Graph → RAG makes it
available to other experts → Observer SDK baseline updates → Fewer false anomalies for that pattern.
11
Expert Reward & Collaboration Framework
To ensure experts actively contribute rather than merely tolerate observation, the system treats
expert judgment as a distinct asset — "Proof of Expertise" — and provides
structured incentives across three tiers.
Tier 1 — Intrinsic & Utility Rewards
Things that make the expert's daily job easier right now.
Mechanism
How It Works
Value to Expert
Personalised Copilot
SDK learns from the expert to build a personalised AI assistant that pre-flags cases matching
their patterns
Less grunt work, more time for complex cases
Compliance Automation
"Why?" button annotations automatically formatted into required audit documentation
Eliminated duplicate administrative writing
Skill Analytics Dashboard
Private metrics showing decision speed, accuracy, and unique strengths vs. baseline
Self-improvement and professional validation
Tier 2 — Extrinsic Recognition
Recognising expertise publicly within the professional community.
Mechanism
How It Works
Value to Expert
"Golden Rule" Leaderboard
Quality-based gamification: experts gain standing when their overrides prove correct over time
Peer recognition and competitive motivation
Named Patterns
Unique heuristics are codified and named after the expert (e.g., "The Chen Technique")
Professional legacy and institutional credit
Expert Review Boards
Top contributors are elevated to a paid review board that resolves low-confidence AI disputes
Authority and paid advisory role
Tier 3 — Financial & Career Compensation
Direct material benefits for contributing IP to the firm.
Mechanism
How It Works
Value to Expert
Knowledge Royalty Pool
A percentage of efficiency gains from AI models trained on expert data is distributed as a bonus
pool
Direct financial reward tied to knowledge impact
AI Training Time Codes
"Why?" button and SIA session time officially coded as "Strategic AI Training" (billable hours)
Knowledge sharing is recognised as productive work, not overhead
Promotion Pillar
Tacit knowledge contribution becomes a formal KPI for reaching Senior/Principal levels
Career advancement tied to institutional contribution
Anti-gaming safeguard: Rewards are tied to outcomes (override
correctness validated over time) and novelty (reasoning is different from what
the baseline AI already knows), preventing performative annotation spam.
12
IT Deployment Architecture
The Observer SDK is deployed to expert workstations through one of four channels, chosen based on
the client's architecture and security constraints.
Deployment Options
Option
Method
Best For
Integration Effort
A
NPM Package (npm install @tacit/observer-sdk)
Modern SPAs (React, Vue, Angular, Next.js)
~2 hours engineering
B
CDN Script Tag (<script src="cdn.tacit.ai/...">)
Legacy server-rendered portals (Java, .NET, PHP)
~1 hour — copy/paste
C
Tag Manager (GTM / Tealium)
When marketing/analytics team controls third-party scripts
The browser extension model is the fastest, lowest-friction deployment path.
IT pushes a private Chrome/Edge extension to expert laptops via corporate MDM (Intune, Jamf).
No source code modifications to the host application are required.
1. IT Pushes ExtensionSilent install via Group Policy / MDM to
expert laptops.
Secure WebSockets (WSS) over port 443, auto-fallback to HTTPS POST
Bandwidth
<2KB per minute. Events batched locally (IndexedDB) and flushed every 5s or 50 events
Data Residency
Regional ingest URLs (e.g., wss://eu-west-2.ingest.tacit.ai) — data never leaves
jurisdiction
PII
All hashing and redaction occurs in the browser before data reaches the network
Key advantage of Option D: Works with any web-based enterprise tool —
Salesforce, Guidewire, Duck Creek, Pega, custom portals — without requiring vendor cooperation
or source code access. Deployment takes days, not months.