📡 MORPHEUS INTELLIGENCE DASHBOARD
🎯 THE THESIS ADVANTAGE
// Academic rigor meets plumbing revenue. Nobody else has this.
$2.1M+
Est. Revenue Impact
🧬 What Makes This Different
Every plumbing company in Kansas City uses the same tools: ServiceTitan for CRM, Google Ads for leads, Yelp for reviews. They all see the same data.
BSP has something none of them have: a custom-built natural language processing engine trained on real review data, with a proprietary dictionary of 3,500+ sentiment terms across 38 linguistic categories. This is not off-the-shelf software. This is a thesis-grade classification system being rebuilt inside Nexus AI to detect fake reviews, score lead quality, optimize ad copy, predict customer behavior, and identify competitive vulnerabilities.
The competitors are looking at stars. BSP is reading the language patterns behind the stars. That is the moat.
📖 DOVE DICTIONARY
// 38 linguistic categories, 3,500+ terms. The DNA of sentiment detection.
📚 All 38 Linguistic Categories
Each category captures a dimension of human language that reveals genuine vs fabricated sentiment.
😡 Anger
😨 Anxiety
😔 Sadness
😃 Positive Emotion
😠 Negative Emotion
🧐 Certainty
🤔 Tentative
⚖️ Causation
💡 Insight
🚫 Negation
💬 Social Words
👪 Family
🤝 Friends
👤 Self-Reference
👥 Other-Reference
📝 Articles
🔗 Prepositions
✨ Adverbs
📏 Quantifiers
🔢 Numbers
💤 Filler Words
🤓 Cognitive Process
👀 Perceptual
💪 Body/Health
🏢 Work
🏆 Achievement
🏠 Home
💰 Money
✝️ Religion
☠️ Death
🔥 Swear Words
⏰ Time
🌍 Space
🎮 Leisure
⤴️ Motion
🔊 Assent
🚫 Dissent
❓ Interrogative
🔍 How Categories Detect Fakes
✅ Real Review Patterns
High self-reference ("I", "my", "we")
Specific perceptual details ("the smell", "I saw")
Time markers ("last Tuesday", "within an hour")
Home/family context ("our bathroom", "kids' room")
Moderate certainty (not absolute claims)
Body/health references ("headache from the leak")
❌ Fake Review Patterns
Low self-reference (generic language)
Excessive positive emotion (over-enthusiasm)
High certainty ("absolutely the best ever")
No perceptual details (no sensory language)
Missing time markers (vague timing)
High achievement words ("excellent", "outstanding")
🤖 NAIVE BAYES CLASSIFIER
// Probabilistic text classification. Trained on real data. Rebuilt for plumbing.
⚙️ How It Works
📝
INPUT
Raw review text
or customer comment
🧬
TOKENIZE
Break into words
Match 38 categories
📊
SCORE
Calculate posterior
probabilities
🎯
CLASSIFY
Sentiment class
+ confidence score
🔧 Plumbing Retraining
The original classifier was trained on Sprint Facebook comments. For BSP, we are retraining on three datasets: (1) BSP's own Google reviews (394+ reviews with star ratings as labels), (2) competitor reviews from Google and Yelp across KC, and (3) ServiceTitan job notes and customer feedback.
# Naive Bayes Classification
P(positive | review) = P(review | positive) * P(positive) / P(review)
# For each word: look up in DOVE Dictionary (38 categories)
# Calculate category frequencies, compare against trained distributions
# Output: class with highest posterior probability
# Training accuracy on Sprint corpus: 87.3%
# Target accuracy on plumbing reviews: 90%+
📚 125,959 COMMENT CORPUS
// Sprint Facebook analysis at scale. The training ground for BSP intelligence.
📊 Corpus Statistics
87.3%
Classification Accuracy
🔄 How This Applies to BSP
The Sprint Facebook corpus proved the methodology works at scale: 125,959 comments classified by sentiment with 87.3% accuracy. For BSP, the same pipeline monitors: Google reviews, Yelp reviews (BSP and competitors), Facebook comments, ServiceTitan feedback, and Daniel AI call transcripts. Instead of just classifying positive/negative, the plumbing version detects: urgency level, service type mentioned, price sensitivity, competitor mentions, and fake review indicators.
⭐ YELP DATASET
// Cross-platform review analysis. Training data for competitive intelligence.
⭐ Cross-Platform Training
📊 Training Data
Yelp reviews with star ratings provide labeled training data. A 5-star review where the language says "horrible" is likely fake. A 1-star review with specific details is genuine frustration.
🔍 Competitive Analysis
BSP competitors on Yelp can be analyzed for review authenticity. If a competitor has 200 reviews but 40% show fake patterns, that intelligence is actionable.
💻 R CODE TO PYTHON
// Porting the thesis analysis pipeline into Nexus AI
🔄 Migration Pipeline
📜
Original: R + LIWC
Thesis used R with the LIWC framework. Ran on local machine. Single-use analysis.
🔄
Port: Python + NLTK + scikit-learn
Rewriting in Python for Nexus integration. NLTK tokenization, scikit-learn Naive Bayes, pandas processing.
🧬
Enhance: DOVE Dictionary as JSON
All 3,500+ terms stored in PostgreSQL on the VM. Fast lookup. Plumbing terms added.
🚀
Deploy: Nexus API Endpoint
REST API: POST /api/thesis/analyze. Returns sentiment, confidence, category breakdown, fake probability.
💻 API Design
# POST /api/thesis/analyze
# Request:
{
"text": "Great service! They fixed our sewer line fast.",
"source": "google_review",
"competitor": false
}
# Response:
{
"sentiment": "positive",
"confidence": 0.91,
"fake_probability": 0.12,
"categories": {
"positive_emotion": 0.34,
"self_reference": 0.22,
"perceptual": 0.18,
"time": 0.11
},
"service_type_detected": "sewer",
"urgency_score": 0.65
}
💰 REVENUE RANKED APPLICATIONS
// All 10 thesis applications, ranked by estimated revenue impact
#1
🚨 Fake Review Detection
$500K+ Impact
Detecting and flagging fake competitor reviews that steal BSP calls. Every removed fake review shifts call volume. At $667 avg job value, removing fakes that redirect 750 calls/year is worth $500K+.
#2
💸 Estimate Follow-Up Priority
$596K Pipeline
$596K in open estimates. The classifier analyzes inquiry language, job notes, and communication patterns to predict which convert. Follow up on the top 20% first. Urgency markers and family references correlate with higher close rates.
#3
📧 Lead Quality Scoring
$322K Campaign
The $322K email campaign sends to 4,832 contacts. The classifier scores each contact's previous communications to rank lead quality. Hot leads get personalized outreach first. This increases the reactivation rate by prioritizing highest-probability converters.
#4
🤖 Daniel AI Call Quality
$2,876/Reactivation
The classifier analyzes call transcripts in real-time to detect sentiment shifts. If urgency markers spike, Daniel pushes for booking. Each successful reactivation averages $2,876 in job revenue.
#5
📝 Ad Copy Optimization
40,847 Keywords
The classifier analyzes which ad copy language patterns generate highest CTR. Which of the 38 categories drive clicks in plumbing ads? Data reveals "emergency" + "family" + "$89" outperforms "professional" + "quality" + "call now".
#6
📰 Blog Content Scoring
576K Monthly Searches
576K monthly sewer searches in KC. The classifier scores blog posts against patterns in top-ranking competitor content. Which categories drive engagement? Data-driven content optimization instead of gut feel.
#7
💬 Customer Sentiment Monitoring
Churn Prevention
Monitor every touchpoint for sentiment shifts. A customer going from positive to negative language is a churn risk. Flag for priority follow-up before they leave a bad review. One saved $6,800 superfan pays for the system.
#8
🔍 Competitor Social Intelligence
Market Positioning
Monitor competitor review sentiment trends monthly. Rising negative sentiment about "wait time"? BSP targets their zip codes with "same-day service" messaging. Real-time competitive intelligence from language analysis.
#9
📧 Email Response Prediction
Open Rate Optimization
Score subject lines before sending. "Your plumbing system is due" (certainty + home) vs "Is your plumbing ready?" (interrogative + tentative). Test results feed back into the model. Data-driven subject line optimization.
#10
⭐ Superfan Identification
Referral Revenue
Identify potential superfans from their first 2 interactions. Current superfans (156 customers) share markers: high home/family language, specific perceptual details, gratitude expressions. Tag new customers who match early, earn referrals. Each superfan = $6,800 LTV + 2-3 referrals.
🔧 DETAILED APPLICATION BREAKDOWNS
// How each application works, what feeds it, and expected ROI
🚨 App 1: Fake Review Detection (Deep Dive)
How it works: Classifier runs against every competitor review. Each gets scored across 38 categories. Reviews matching fake patterns (excessive positive + low self-reference + high certainty + missing perceptual) flagged with fake probability.
Pipeline: Pull competitor reviews nightly. Score against DOVE Dictionary. Run Naive Bayes (genuine vs fake). Flag reviews with fake probability > 70%. Generate evidence report. Robert reviews and reports to Google.
ROI: Removing 20 fake competitor reviews shifts ~100+ calls/year to BSP. At $667 avg = $66,700 recovered. Ceiling much higher with systematic monitoring.
💸 App 2: Estimate Follow-Up Priority (Deep Dive)
How it works: Classifier scores each estimate's inquiry language for urgency, commitment signals, and objection markers.
Scoring: High urgency + certainty + family = HOT (follow up 24hrs). Medium urgency + money concerns = WARM (financing options). Low urgency + dissent = COLD (nurture sequence).
ROI: If priority scoring increases close rate from 35% to 40% on $596K in estimates = $29,800 additional revenue. With Ashton's sewer data ($262K from 150 points), the signal is clear.
📧 App 3: Lead Quality Scoring
How it works: Score each contact's historical communications before the email campaign sends. Customers with specific, detailed language (perceptual words, time markers, home references) are higher quality.
ROI: Top 20% get premium outreach, bottom 20% get basic templates. Expected 18% reactivation on top tier vs 5% on bottom. Net revenue improvement of $40K-$80K over flat targeting.
🤖 App 4: Daniel AI Call Quality
Mechanism: Real-time transcript analysis. When "I need to think about it" (tentative + cognitive) detected, Daniel switches to reassurance. When urgency markers ("flooding", "emergency") appear, Daniel pushes for immediate booking.
ROI: Improving booking rate from 7 to 10+ per month = $28,760+ additional monthly revenue.
📝 App 5: Ad Copy Optimization
Mechanism: Score RSA headlines against 38 categories. Track which combinations produce highest CTR. Feed winning patterns into new copy generation.
ROI: 10-15% CTR improvement = lower CPC = more leads at same budget.
📰 App 6: Blog Content Scoring
Mechanism: Score blog drafts against top-ranking competitor content linguistic patterns. If competitors use perceptual language and your draft is heavy on achievement language, adjust.
ROI: Better ranking = more organic traffic = reduced ad spend dependency.
💬 App 7: Customer Sentiment Monitoring
Mechanism: Build a sentiment timeline per customer across all touchpoints. Negative trajectory = auto-flag for personal outreach before churn.
ROI: Preventing 10 losses/year at $6,800 avg superfan value = $68,000.
🔍 Apps 8-10: Competitor Intel, Email Prediction, Superfan ID
Competitor Social Intel: Monthly sentiment trend monitoring. Rising negative sentiment = targeting opportunity.
Email Response Prediction: Score subject lines pre-send using category analysis. A/B test results feed back into the model.
Superfan ID: Match first 2 interactions against known superfan language patterns. 20 new superfans/year = $136,000+ in lifetime value.
🧬 THE THESIS IS NOT ACADEMIC. IT IS A REVENUE ENGINE.
Nobody in Kansas City plumbing has custom NLP. Nobody has a proprietary sentiment dictionary. Nobody is analyzing competitor reviews at the linguistic level. Nobody is scoring leads based on language patterns.
BSP does. That is the moat. The thesis built the foundation. Nexus makes it operational.
10 applications. $2.1M+ estimated total impact. Every one builds on the same core: the DOVE Dictionary + Naive Bayes Classifier.
🔎 HOW FAKE REVIEW DETECTION WORKS
// 6 engines powered by Robert's thesis dictionaries. From review to evidence in seconds.
STEP 1: COLLECT
📡 Google Places API Competitor Scraping
TARGET COMPETITORS
Cororum Plumbing
Dick Ray Plumbing
Inception Plumbing
A-1 Sewer & Septic
AB May
DATA PER REVIEW
• Full review text + star rating
• Exact post date and time
• Reviewer display name
• Reviewer total review history
• Other businesses reviewed by same account
STEP 2: ANALYZE
ENGINE 01
Sentiment Uniformity
Weight: 15%
- Bing Liu lexicon: 6,789 words
- Real reviews show mixed sentiment
- Fake reviews = 100% positive, zero negatives
SUSPICIOUS: 0 negative words in 200-word review
THESIS WEAPON
ENGINE 02
Linguistic Authenticity
Weight: 25% (highest)
- DOVE Dictionary: 3,500+ words, 38 categories
- Checks hedging, hesitators, first person, discourse
- Real reviews use 5+ categories. Fake = 0-1.
SUSPICIOUS: Only "achievement" category detected
ENGINE 03
Temporal Velocity
Weight: 20%
- Real pattern: 2-3 reviews per week, spread out
- Fake pattern: 20 reviews in 48 hours
- Detects burst campaigns and seasonal spikes
SUSPICIOUS: 15 five-star reviews in 36 hours
ENGINE 04
Reviewer Profile
Weight: 20%
- One-hit wonder accounts (single review ever)
- Cross-competitor reviewers (reviewed 3+ competitors)
- Account age vs review count mismatch
SUSPICIOUS: Account created 2 days before review
ENGINE 05
Text Similarity
Weight: 10%
- TF-IDF vectorization + cosine similarity
- 80%+ match = copy/paste or template
- Catches review farms using boilerplate text
SUSPICIOUS: 87% match with 3 other reviews
ENGINE 06
AI Content Detection
Weight: 10%
- Sentence length uniformity (AI = consistent)
- Vocabulary diversity score (AI = limited range)
- Formal transition markers ("Furthermore", "Moreover")
SUSPICIOUS: Perplexity score 12.3 (human avg: 45+)
STEP 3: SCORE
📊 Weighted Composite Scoring
Each engine contributes its weighted score to produce a single fake probability (0-100).
FAKE_SCORE = (Sentiment_Uniformity * 0.15) + (Linguistic_Auth * 0.25)
+ (Temporal_Velocity * 0.20) + (Reviewer_Profile * 0.20)
+ (Text_Similarity * 0.10) + (AI_Content * 0.10)
AUTHENTIC
SUSPICIOUS
LIKELY FAKE
HIGH CONF.
0395979100
STEP 4: WEAPONIZE
📋
Evidence Package
Auto-generate screenshot, profile analysis, temporal chart, and similarity matches for each flagged review
🚩
Report to Google
Submit flagged reviews through GBP with full evidence documentation and pattern analysis
📩
Send to Evelyn
Package evidence for Google rep Evelyn with data-backed analysis and removal request
📊
Dashboard Update
Live results at the Morpheus Competitor Review Intelligence dashboard
CURRENT RESULTS
📊 Competitor Review Authenticity Scores
| Company | Avg Score | Status | Notes |
|---|
| Bright Side Plumbing | 12.2 | CLEAN | Lowest score = most authentic. 394+ reviews. |
| Dick Ray Plumbing | 17.5 | CLEAN | Authentic review profile |
| A-1 Sewer & Septic | 18.5 | CLEAN | Authentic review profile |
| Cororum Plumbing | 19.6 | CLEAN | Authentic review profile |
| AB May | 21.2 | 2 BORDERLINE | Two reviews flagged for temporal clustering |
| Inception Plumbing | 42.4 | 3 SUSPICIOUS | Low linguistic diversity + temporal burst pattern |
BSP scored lowest (12.2) = most authentic review profile in KC plumbing.
THE MATH
💰 Revenue Impact Analysis
4.9⭐
BSP Rating (392 REAL)
750+
Misdirected Calls/Year
Fake reviews steal calls from BSP by inflating competitor ratings. Each removed fake review adjusts the competitor's rating downward, making BSP's legitimate 4.9-star rating with 392 real reviews the clear winner. At BSP's average job value of $667, shifting misdirected calls back to BSP represents $500K+ in annual recovered revenue.
TO GO LIVE
🚀
System Built. Ready to Deploy.
✅ Step 1
Enable Google Places API in GCP Console
✅ Step 2
Configure API key and competitor Place IDs
✅ Step 3
Activate nightly scraping + analysis pipeline