Scaling Connoisseurship

1. Dual Mandate: Investment Merit + Curatorial Coherence

Egon refuses the false choice between investment returns and aesthetic coherence. Every recommendation balances quantitative investment grading (A+ to C-, detailed in Investment Science whitepaper Section 4) with aesthetic alignment scoring (1-10 scale). An A+ investment with 4/10 aesthetic fit generates INVESTMENT_ONLY verdict—quality artist, wrong for your collection. An artist perfectly aligned (9/10) but with weak fundamentals (C grade) triggers WAIT_FOR_VALIDATION—love the fit, unfavorable market timing.

This whitepaper details Egon's aesthetic intelligence system: how we discover preferences collectors cannot articulate (Section 2), structure them for evolution tracking (Section 3), learn from implicit behavior signals (Section 4), enable semantic matching via vector embeddings (Section 5), calculate alignment scores (Section 6), and assess curatorial quality beyond ROI (Section 7).

2. Conversational Onboarding: Discovering Unarticulated Preferences

Most collectors cannot articulate their aesthetic preferences systematically. They know they like "expressive paintings" or "bold colors" but struggle to define collecting themes, identify stylistic priorities, or explain why certain artists resonate. Egon's conversational onboarding solves this through AI-driven discovery.

Three-Stage Discovery Process

Figure 1: Conversational onboarding discovers comprehensive aesthetic profiles through intelligent dialogue (8 exchanges total).

Stage 1: Aesthetic Discovery (4-5 exchanges)

Initial Prompt: "Tell me about a piece that has genuinely moved you - what was it, and what specifically about it caught your attention? Or, if you already have a list of favorite artists, feel free to drop them in the chat."

AI Intelligence: Claude extracts preferences from natural conversation:

Confidence Calibration: AI progressively adjusts theme_clarity based on exchange count, with confidence increasing from initial exchanges (low confidence, directional) through mid-exchanges (emerging patterns) to final exchanges (high confidence, well-defined understanding). This graduated calibration prevents premature synthesis while ensuring profile completeness within reasonable interaction time.

Edge Case Handling: When users provide minimal responses ("I don't know", "Whatever you think"), the system employs forced advancement after exchange limits to prevent endless onboarding loops while ensuring adequate signal collection. Response richness analysis adjusts phase progression accordingly—users providing rich detail may complete phases faster, while minimal responders still advance after exchange limits to maintain UX momentum.

Stage 2: Investment Discovery (2-3 exchanges)

Transition: "Now let's talk about your investment approach. What's your budget range, and are you thinking short-term (3-5 years) or building a long-term legacy collection?"

Extracted Parameters:

• Initial budget, annual acquisition budget, budget flexibility
• Risk tolerance (conservative/balanced/aggressive)
• Investment timeline (short-term/medium-term/long-term/legacy)
• Primary objective (appreciation/passion/dual/institutional legacy)
• Liquidity needs (high/moderate/low)
• Portfolio allocation targets by tier (Blue-chip/Value/Growth/Discovery)

Stage 3: AI Synthesis & Profile Generation

After 6-8 exchanges, Claude synthesizes a comprehensive profile including:

Implementation note: Onboarding conversation managed with phase-based progress tracking, forced advancement after exchange limits (prevents endless aesthetic discussions), and confidence-calibrated synthesis.

3. Dual-Table Architecture: Core Identity + Granular Preferences

Egon uses a sophisticated dual-table system separating stable core identity (UserProfileV2) from evolving fine-grained preferences (AestheticPreference). This architecture enables:

• Preference evolution tracking: Historical record of how tastes change (PreferenceEvolution table)
• Source attribution: Preferences from onboarding vs learned from user behavior
• Confidence scoring: High-confidence explicit preferences vs inferred signals
• Weight adjustment: Preferences strengthen/weaken based on user actions
• Temporal filtering: Query active preferences vs full historical record

Core Identity Table

# Stable identity data (rarely changes)
class UserProfile:
    user_id: int

    # Personal & Financial Context
    identity_fields: dict      # Demographics, financial capacity

    # Investment Strategy
    strategy_fields: dict      # Risk tolerance, timeline, allocation targets

    # Collection Vision (AI-Synthesized)
    vision_narrative: str      # Thematic direction synthesized from onboarding
    vision_confidence: float   # Synthesis quality score
    vision_keywords: list      # Extracted key themes

    # Behavioral Profile
    behavioral_fields: dict    # Experience level, decision style, pace

    # Advisory Configuration
    flexibility_level: str     # How open to exploration
    advisor_preference: str    # Communication style preference

    # Multi-Dimensional Embeddings
    embeddings: dict           # Specialized vectors for semantic matching

Granular Preferences Table

# Fine-grained, evolving preferences (many records per user)
class AestheticPreference:
    user_id: int

    # Preference Data
    preference_type: str       # Category (artist, movement, medium, theme, etc.)
    preference_value: str      # Specific value

    # Scoring Metadata
    weight: float              # Strength of preference (weighted scoring)
    source: str                # Origin of preference (onboarding, implicit learning, etc.)
    confidence: float          # Certainty level (explicit vs inferred)

    # Temporal Tracking
    status_fields: dict        # Active/historical status, reinforcement tracking

Preference Type Taxonomy

4. Preference Evolution: Implicit Learning from User Actions

Egon continuously refines aesthetic profiles through implicit learning—observing user behavior patterns without requiring explicit feedback. Four primary learning mechanisms:

1. Watchlist Additions

# User adds artist to watchlist
# System infers multiple preferences from single action:

new_preferences = [
    {
        'type': 'artist',
        'value': 'Artist Name',
        'weight': MODERATE_SIGNAL,      # Explicit action = moderate confidence
        'source': 'watchlist_action',
        'discovered_via': 'implicit_learning'
    },
    # Inferred from artist's characteristics:
    {
        'type': 'investment_priority',
        'value': 'Tier engagement pattern',
        'weight': INFERRED_SIGNAL,      # Lower confidence (inferred)
        'context': 'Derived from artist tier classification'
    },
    {
        'type': 'movement',
        'value': 'Artist movement',
        'weight': INFERRED_SIGNAL,
        'context': 'Inferred from artist addition'
    }
]

2. Collection Acquisitions

# User acquires artwork
# Strongest signal type (actual purchase = high commitment)

new_preferences = [
    {
        'type': 'artist',
        'value': 'Artist Name',
        'weight': STRONG_SIGNAL,        # Purchase = very high confidence
        'source': 'collection_acquisition',
        'is_primary': True              # Mark as core preference
    },
    {
        'type': 'style_descriptor',
        'value': 'Artist style characteristics',
        'weight': MODERATE_HIGH_SIGNAL,
        'context': 'Derived from acquisition'
    },
    {
        'type': 'acquisition_criterion',
        'value': 'Price point validation',
        'weight': MODERATE_HIGH_SIGNAL,
        'context': 'Inferred budget tolerance'
    }
]

# Also UPDATE existing related preferences:
# Reinforce movement/style preferences that align with acquisition
# Update temporal tracking for related preferences

3. Analysis Corrections (Source: analysis_correction)

When users disagree with Egon's verdict, we learn from the mismatch:

4. Interaction Patterns (Source: engagement_analysis)

System analyzes aggregate behavior patterns monthly:

• Analysis request patterns: Frequent analyses of Discovery-tier artists → INCREASE Discovery preference weight
• Watchlist composition: Strong representation of Color Field painters → REINFORCE Color Field movement preference
• Search behavior: Frequent semantic searches for "minimalist abstraction" → ADD/STRENGTHEN theme_element preferences
• Time-based decay: Preferences without recent reinforcement → DECREASE weight gradually over time

Preference Evolution Tracking

A dedicated evolution tracking system maintains historical records of preference changes:

• Change types: add, update, strengthen, weaken, remove
• Change reasons: implicit learning, explicit update, time decay
• Trigger attribution: Links changes to specific user actions
• Context preservation: Rich explanations for future analysis

This enables sophisticated preference analytics—understanding not just what a collector likes today, but how their taste has evolved over time and which experiences shaped those changes.

5. Multi-Dimensional Vector Embeddings

Egon generates 3 specialized high-dimensional embeddings per user profile, enabling weighted multi-dimensional similarity queries impossible with keyword-based systems. This mirrors the multi-embedding architecture for artists detailed in Data Universe whitepaper Section 5.

User Profile Embeddings

Update Cadence and Triggers

Embedding Generation Strategy: Smart field change detection prevents unnecessary re-embedding. Embeddings regenerate via asynchronous batch processing triggered by:

• collection_vision_embedding: Regenerates when core vision fields change
• aesthetic_taste_embedding: Periodic batch processing triggered by preference evolution metrics—only when aggregate preference drift exceeds threshold, not on every individual change
• investment_profile_embedding: Regenerates when investment parameters change
• Initial generation: All embeddings created immediately after onboarding completion

Cost Efficiency: Event-driven regeneration with threshold-based batching minimizes API costs while maintaining freshness.

Weighted Multi-Dimensional Queries in Production

Embeddings enable sophisticated similarity matching with custom weighting across Egon's tools:

# Example 1: Semantic artist search
# User: "Find artists like Rothko but with more geometric structure"

query_vector = weighted_combination(
    user.aesthetic_taste_embedding,      # Primary taste preference
    user.collection_vision_embedding     # Thematic coherence
)

similar_artists = vector_search(
    query_vector=query_vector,
    target_dimensions=['aesthetic', 'institutional'],
    filters={'tier': ['Growth', 'Discovery']},
    limit=15
)

# Result: Artists aesthetically similar to user's preferences
# with institutional validation, filtered by investment tier
# Semantic understanding enables natural language queries

# Example 2: Thematic coherence scoring
# Used in every artist analysis

similarity = cosine_similarity(
    artist.aesthetic_embedding,
    user.collection_vision_embedding
)

thematic_coherence_score = normalize_to_scale(similarity)

# Enables semantic matching between artist characteristics
# and collector's narrative vision

# Example 3: Portfolio gap analysis
# Identify artists that fill thematic/stylistic gaps

# Aggregate current collection's aesthetic profile
collection_profile = aggregate_collection_embeddings(user.collection)

# Find thematically coherent but stylistically novel artists
gap_filling_artists = vector_search(
    query_vector=weighted_combination(vision, taste),
    filters={
        'exclude_movements': current_movements,  # Find NEW movements
        'min_quality_threshold': QUALITY_FLOOR
    },
    limit=20
)

# Result: Artists that strengthen collection coherence
# while filling identified gaps

Why Multi-Dimensional Embeddings Matter

Separating vision, taste, and investment embeddings enables queries impossible with a single unified embedding:

• Vision-focused search: "Find artists that fit my collection's phenomenological exploration theme regardless of my current taste preferences" → Query primarily on collection_vision_embedding, weight aesthetic_taste_embedding low
• Taste-focused search: "Show me artists I'd personally love even if they don't fit my stated collection theme" → Query primarily on aesthetic_taste_embedding (exploratory discovery)
• Investment-aligned search: "Find conservative Blue-chip opportunities matching my risk profile" → Filter on investment_profile_embedding similarity before aesthetic matching
• Balanced search: Default discovery mode combines all 3 dimensions with custom weights based on user's primary_objective (appreciation-focused: 0.3 vision + 0.3 taste + 0.4 investment; passion-focused: 0.4 vision + 0.5 taste + 0.1 investment)

Collector Clustering (Future Enhancement)

Multi-dimensional embeddings enable sophisticated collector segmentation for peer insights:

• Taste-based clusters: Group collectors by aesthetic_taste_embedding similarity → identify shared preferences, collaborative filtering recommendations ("Collectors with taste similar to yours also love Carmen Herrera")
• Vision-based clusters: Group by collection_vision_embedding → collectors with similar thematic narratives, institutional trajectories ("Your Color Field → Minimalism narrative aligns with museum-quality collectors in Cluster A")
• Investment-based clusters: Group by investment_profile_embedding → peer portfolio performance benchmarking ("Your Sharpe ratio: 2.3 vs peer cluster median: 1.8—outperforming similar-risk collectors")
• Multi-cluster insights: "You share aesthetic taste with Cluster A (Color Field collectors) but investment approach with Cluster B (aggressive Discovery-tier investors) → unique positioning for undervalued Color Field emerging artists"

6. Aesthetic Alignment Scoring: Balancing Investment & Coherence

Every artist analysis includes dual scoring: Investment Grade (A+ to C-) measuring market fundamentals + Aesthetic Alignment (1-10 scale) measuring collection fit. Both scores feed into the 15-category verdict system detailed in Investment Science whitepaper Section 4.

Alignment Scoring Formula (Hierarchical Weighting)

Alignment scores are calculated as a weighted combination of three components with proprietary hierarchical weighting:

• Direct matching score — Explicit preference matches (primary weight)
• Thematic coherence score — Theme/narrative fit (secondary weight)
• Contextual fit score — Portfolio gap filling (tertiary weight)

Weighting Hierarchy Methodology

The hierarchical weighting reflects preference signal strength observed in art advisory practice:

• Primary Weight - Direct Matching: Explicit preferences (artist, movement, medium) are the strongest signals of collector intent. When a user states "I love Rothko" during onboarding or adds an artist to their watchlist, this represents clear commitment—the primary component of alignment scoring reflects these explicit signals.
• Secondary Weight - Thematic Coherence: Collection theme provides important validation of aesthetic fit. An artist may not match explicit preferences but could strongly align with the collector's broader thematic vision (e.g., "phenomenological exploration through reductive form")—this deserves substantial but not dominant weight.
• Tertiary Weight - Contextual Fit: Portfolio gap-filling is valuable but contextual. An artist filling a demographic gap (e.g., first woman artist in collection) provides curatorial benefit but shouldn't override aesthetic misalignment—hence receives the lowest weight in the hierarchy.

Similarly, dynamic thresholds (varying by collector flexibility: focused collectors require higher alignment scores, exploratory collectors accept lower thresholds) and progressive confidence calibration across onboarding exchanges were designed to balance UX with profile accuracy: prevent premature synthesis (early exchanges = lower confidence) while avoiding diminishing returns from extended questioning (exchange limits enforce completion). Preference weight adjustments reflect commitment intensity hierarchy—actual acquisitions signal stronger preference than watchlist additions, which in turn outweigh passive browsing.

Methodological note: These are principled design decisions informed by art advisory practice and UX research, not empirically derived from large-scale data. Quantitative validation is planned as the platform matures and sufficient user interaction data accumulates (target: 500+ collectors, 5,000+ preference signals for statistically robust optimization).

Component 1: Direct Matching (Primary component)

Exact matches against stored preference records with weight-adjusted scoring:

# Example: Analyzing artist for user with specific movement preferences

direct_matches = [
    {'type': 'artist', 'match': 'exact'},      # User has explicit artist preference
    {'type': 'movement', 'match': 'exact'},    # Movement alignment
    {'type': 'medium', 'match': 'exact'},      # Medium preference match
    {'type': 'theme_element', 'match': 'partial'},  # Partial thematic overlap
]

# Calculate weighted average using proprietary weighting
# Exact matches weighted higher than partial matches
direct_matching_score = calculate_weighted_match_score(direct_matches)

# Result: Normalized 1-10 score based on match quality and preference weights

Component 2: Thematic Coherence (Secondary component)

Semantic similarity between artist's work and user's collection theme using vector embeddings:

# Compare artist and user embeddings
similarity = cosine_similarity(
    artist.aesthetic_embedding,
    user.collection_vision_embedding
)

# Convert to 10-point scale
thematic_coherence_score = normalize_to_scale(similarity)

# Result: High similarity indicates strong thematic fit
# e.g., Color Field artist vs "exploring tension in post-war abstraction"

Component 3: Contextual Fit (Tertiary component)

Portfolio gap analysis—does this artist fill a strategic gap or create redundancy?

# Evaluate multiple contextual factors

contextual_factors = []

# Movement diversity analysis
if fills_movement_gap(artist, user_collection):
    contextual_factors.append({'reason': 'Building movement depth', 'score': HIGH})
elif creates_movement_overweight(artist, user_collection):
    contextual_factors.append({'reason': 'Movement overweight', 'score': REDUCED})

# Demographic diversity analysis
if fills_diversity_gap(artist, user_collection):
    contextual_factors.append({'reason': 'Strengthens diversity', 'score': HIGH})

# Tier allocation analysis
if fills_tier_gap(artist, user_allocation_targets):
    contextual_factors.append({'reason': 'Fills allocation gap', 'score': HIGH})

# Aggregate contextual factors
contextual_fit_score = calculate_contextual_average(contextual_factors)

Dynamic Threshold System

Alignment thresholds adapt to user's aesthetic flexibility setting:

Borderline Case Handling

Scores near thresholds receive special treatment in verdict logic:

• Below threshold: Triggers INVESTMENT_ONLY verdict if investment grade is strong, preventing recommendations of quality artists that don't quite fit. However, if investment grade is mediocre, generates ADJACENT_STYLE verdict—acknowledging partial aesthetic alignment and encouraging exploration.
• Just above threshold: Enters BUY consideration logic but receives confidence penalty in final verdict. Threshold-adjacent alignments with strong investment become BUY_OPPORTUNISTICALLY (not STRONG_BUY), signaling "good fit but not perfect match."
• Focused user below strict threshold: Generates WAIT_FOR_VALIDATION verdict even with strong fundamentals—focused collectors shouldn't compromise on aesthetic fit.
• Exactly at threshold: Treated as passing the threshold, preventing threshold boundary inconsistency.

This nuanced handling prevents binary threshold artifacts while maintaining clear decision boundaries. The ADJACENT_STYLE verdict category specifically addresses partial matches—scores showing thematic proximity but not full alignment.

7. Curatorial Collection Analysis: Beyond ROI

Egon's portfolio analysis evaluates not just financial metrics (Sharpe ratio, returns, allocation) but curatorial quality—the strength of the collection as a cohesive artistic statement. This reflects our philosophy: we build art collectors, not just investment portfolios.

Four Curatorial Dimensions

1. Thematic Coherence (1-10 scale)

How well do the artworks relate to each other and the stated collection theme?

# Calculate theme coherence

# Step 1: Compare each artwork against collection theme
artworks_alignment_scores = []
for artwork in collection:
    similarity = cosine_similarity(
        artwork.aesthetic_embedding,
        user.collection_vision_embedding
    )
    artworks_alignment_scores.append(similarity)

# Step 2: Identify thematic outliers
outliers = identify_low_alignment_artworks(artworks_alignment_scores)

# Step 3: Calculate coherence with consistency bonus
coherence_score = calculate_coherence(
    mean_alignment,     # Base alignment
    alignment_variance  # Reward consistency
)

# Interpretation scale:
# High scores: Museum-quality curation, clear thematic thread
# Moderate scores: Emerging direction, developing coherence
# Low scores: Scattered collection, unclear theme

2. Collection Narrative Strength

Can the collection tell a compelling art historical story?

• Movement diversity: Does the collection explore a movement's evolution (early → mature phases) or show cross-movement dialogue?
• Artist depth: Multiple works by key artists vs single representative pieces
• Medium balance: Coherent medium focus (all painting) vs intentional mixed-media exploration
• Temporal arc: Historical depth (spanning decades) vs contemporary focus
• Institutional validation: % of collection with museum pedigree

3. Strategic Gap Analysis (3 Dimensions)

What's missing to strengthen curatorial coherence?

4. Collection Stage Assessment

Evaluate collection's developmental trajectory and institutional trajectory:

Curatorial vs Investment Prioritization

Portfolio analysis recommendations balance both dimensions based on user's primary_objective:

Portfolio Analysis Output Structure

# Comprehensive portfolio analysis includes:

{
    'financial_metrics': {
        'total_value': value,
        'total_invested': cost_basis,
        'unrealized_gains': gains,
        'annualized_return': performance,
        'sharpe_ratio': risk_adjusted_return,
        'allocation': allocation_by_tier
    },

    'curatorial_assessment': {
        'coherence_score': thematic_unity_score,
        'narrative_strength': 'Collection narrative description',
        'thematic_strengths': [identified_strengths],
        'collection_stage': stage_classification,
        'institutional_trajectory': validation_assessment,

        'strategic_gaps': [
            {
                'type': 'gap_category',
                'description': 'Gap description',
                'priority': priority_level,
                'suggested_artists': [aligned_recommendations],
                'rationale': 'Why this acquisition strengthens collection'
            }
            # ... additional gaps
        ],

        'next_strategic_priorities': [
            'Prioritized acquisition recommendations',
            'Balancing curatorial and investment considerations'
        ]
    },

    'verdict': 'Comprehensive assessment synthesizing financial performance with curatorial quality, identifying priority actions that strengthen both dimensions.'
}