Ethical Score-Based Intelligence

The Engine:
Vector Interest Profiles

No raw spying. No black box. Just transparent, decaying scores that evolve with you.
User ≠ Label. User = Evolving Vector.

1. The Principle: You Are A Score, Not A Label

Most platforms tag you as "Student" and leave you there. We allow you to be complex. Your profile is made of confidence scores that rise and fall.

  • Explainable

    "You see this because your hiking score is 0.82."

  • Decaying

    If you stop hiking, your hiking score drops naturally.

user_interest_vector.json
"hiking": 0.82 // High
"coworking": 0.74 // Active
"budget_living": 0.66 // Moderate
"luxury_lifestyle": 0.12 // Low

2. The Three Signal Layers

Where do these scores come from? We use 3 explicit sources.

A

Cold Start (Onboarding)

WEIGHT: 40% (Initial)

Explicit data users give us when they join. It sets the baseline but fades over time.

  • Selected City
  • Budget Range
  • Chosen Interests
B

Behavioral (The Truth)

WEIGHT: 40% (Growing)

What users actually do. This is the strongest signal and naturally overtakes the initial profile.

  • Dwell Time (18s+)
  • Communities Joined
  • Events Saved
C

Social Connections

WEIGHT: 20% (Context)

Optional external signals converted into abstract scores (never raw data).

  • LinkedIn (Career)
  • .edu Email (Student)
  • Friend Circles

3. Ethical Scoring Rules

You are in control. The system cleans itself automatically so you never get stuck in an old bubble.

Transparency

"You were added to 'Lisbon Surfers' because you viewed 5 surf events."

Natural Decay (The Anti-Bubble)

Old interests fade. If you stop clicking 'Party', your profile shifts to 'Chill'.

Score_new = Score_old × 0.98 (per week)

Live Decay Simulation

Party InterestWeek 1: 0.90
Party InterestWeek 4: 0.65
Party InterestWeek 8: 0.15
Status: Interest Faded Successfully

4. Gravity Auto-Join System

We don't ask you search. We place you where you belong.

The Formula

(City_Match × 0.4)+
(Interest_Overlap × 0.4)+
(Persona_Rel × 0.2)=
Join_Score> 0.7 = AUTO JOIN

If the score is high enough, you are instantly added to the hub.
Max 3-5 core hubs to prevent overwhelm.

Smart Suggestions

For everything else, we suggest rather than force.

Lisbon HikersConfidence: 82%
Outdoor
Active
AI Reason: "You engaged with 3 outdoor events this week."

5. One Brain. Many Outputs.

Your Interest Vector drives the entire platform experience.

Feed Ranking
Reels Shown
Push Notifs
Event Suggestions
Deal Matching
Service Core

6. Predictive Recommendation Engine

Our original powerhouse. It scans explicit signals (filters, searches) and contextual data (location, budget) to proactively find the perfect housing, visa services, and local perks.

  • Intent-Based Search Ranking
  • Hyper-Local Service Matching
  • Real-Time Inventory (APIs)
Targeting: "Lisbon" + "Student" + "Low Budget"
98% Match
Student Residence - Marques
Verified • €450/mo • Wi-Fi

7. The Intelligence Pipeline

How meaningful signals become magical experiences.

01

Signal Ingestion

The system focuses on implicit data: Time spent on a card (Dwell Time), Community Hub joins, Job saves, and search filters. This builds a "Context Vector" without the user needing to type a single word.

02

Semantic Intent

The NLP Cortex uses Semantic Embeddings (Hugging Face) to understand concepts. It knows that reading about "Blue Card" means "Visa" and "Relocation", connecting dots across different categories.

03

Community Graph

The engine looks at your Community Hub memberships. If you join "Digital Nomads in Bali", it filters housing for "Laptop Friendly" and "High-speed WiFi" automatically.

04

Predictive Ranking

Finally, the LambdaMART Ranker scores every item based on probability. It promotes the hidden gem that perfectly matches your budget and your friend's recent activity.

8. Core Algorithms & Tech Stack

The mathematical engine running underneath.

TensorFlow
Core Intelligence
LightGBM
LambdaMART Ranking
Hugging Face
Semantic Embeddings
NetworkX
Social Graph Engine
FastAPI
High-Performance API
scikit-learn
Predictive Analytics
Redis
Real-Time Cache
FAISS
Vector Retrieval