On-Device
FuturePrivacy-First Personalization
All preference learning happens on-device. User data never leaves the browser. GDPR/CCPA compliant by design.
The Problem with Traditional Personalization
Traditional Approach
- Every click, hover, and scroll sent to server
- Browsing history stored in databases
- Profile built from cross-site tracking
- Data sold to third parties
- Complex consent dialogs required
V3 Approach
- All behavior analysis on-device
- Preference model stays in browser
- No cross-site tracking possible
- User controls their data completely
- Minimal consent needed (no PII collected)
On-Device Learning
The preference model learns from user behavior without sending data anywhere. Stored in IndexedDB, it persists across sessions while remaining fully private.
// On-device preference learning
class LocalPreferenceModel {
private db: IDBDatabase;
private embedder: Pipeline;
async learnFromBehavior(signal: BehaviorSignal) {
// Generate embedding for the interaction
const embedding = await this.embedder(signal.context);
// Update local preference weights
const update = {
category: signal.productCategory,
weight: this.calculateWeight(signal),
embedding: Array.from(embedding.data),
timestamp: Date.now()
};
// Store in IndexedDB (never sent to server)
await this.db.put('preferences', update);
}
async getPersonalizedRanking(products: Product[]) {
const preferences = await this.db.getAll('preferences');
return products.map(p => ({
...p,
relevanceScore: this.computeRelevance(p, preferences)
})).sort((a, b) => b.relevanceScore - a.relevanceScore);
}
// User controls their data
async exportData() {
return await this.db.getAll('preferences');
}
async deleteAllData() {
await this.db.clear('preferences');
}
}Optional: Federated Learning
For users who opt-in, V3 supports federated learning—aggregate insights without sharing individual data. Only model gradients are shared, never raw behavior.
How Federated Learning Works
Local Training
Model trains on your device using your behavior
Gradient Extraction
Only weight updates extracted, not data
Secure Aggregation
Server combines gradients from many users
Global Model Update
Improved model distributed back to devices
Key insight: Server never sees individual behavior—only aggregated, anonymized model improvements. Individual users cannot be identified.
Compliance by Design
GDPR Compliance
- Data Minimization: No PII collected
- Purpose Limitation: Only used for personalization
- Right to Access: exportData() function
- Right to Erasure: deleteAllData() function
- Data Portability: JSON export supported
CCPA Compliance
- No Sale of Data: Data never leaves device
- Right to Know: Full transparency on what's stored
- Right to Delete: One-click data deletion
- No Discrimination: Same experience with/without tracking
User Control
Privacy Dashboard
On-Device Learning
Learn from your behavior locally
Federated Learning
Contribute to collective insights (opt-in)