Realays AI Use Cases: FruitsFace & Dalendar

So far, we’ve explored Edge Intelligence and Web ML technology. However, technology only has meaning when it delivers value to users. At Realays, we’ve incorporated the cutting-edge technologies mentioned earlier into actual services, providing users with new and secure experiences.

1. FruitsFace: Face Analysis Without Privacy Concerns

FruitsFace is an entertainment service that analyzes a user’s face and matches them with the fruit character they resemble most.

Problem Awareness

Facial photos are highly sensitive biometric information. Sending them to a server for analysis always carries the risk of data leakage. Traditional face analysis services require uploading your photo to external servers, where it could potentially be:

• Stored indefinitely without your knowledge
• Accessed by unauthorized parties
• Used for training AI models
• Sold to third parties

This creates a fundamental trust problem between users and services.

Technical Solution

We chose to use TensorFlow.js-based Face Detection and Classification models that run directly in the user’s browser.

Technical Architecture:

// FruitsFace web ML pipeline
async function analyzeFace(imageElement) {
  // 1. Load face detection model (client-side)
  const detector = await faceapi.nets.tinyFaceDetector.loadFromUri("/models");

  // 2. Detect face in image
  const detection = await faceapi.detectSingleFace(imageElement);

  // 3. Extract facial features
  const landmarks = await faceapi.detectFaceLandmarks(imageElement);

  // 4. Match with fruit database
  const fruitMatch = await matchFruit(landmarks);

  return fruitMatch;
}

Key Technologies Used:

• Face Detection: TensorFlow.js with MobileNet architecture
• Feature Extraction: 68-point facial landmark detection
• Classification: Custom-trained fruit matching algorithm
• Performance: Runs at 30 FPS on modern devices

User Value

Privacy Guarantee: Users can enjoy the service with peace of mind, knowing their photos never leave their device. All processing happens locally in the browser—no data is transmitted to our servers.

Instant Response: No server round-trip time means results appear in real-time. Point your camera at your face and see the fruit match instantly—a snappy response that delights users.

Offline Capability: After the initial model load, FruitsFace works completely offline. Perfect for users concerned about data usage or in areas with poor connectivity.

Real-World Impact

Since launch, FruitsFace has:

• Processed over 1 million faces
• Achieved 99.9% privacy compliance
• Maintained sub-100ms inference time
• Received 4.8/5 user satisfaction rating

Users particularly appreciate:

• “Finally, a face app I can trust!”
• “Works on my phone without eating up my data”
• “So fast, feels like magic”

2. Dalendar: Emotion Assistant in My Diary

Dalendar is a smart diary that analyzes diary entries to automatically record daily emotions.

Problem Awareness

Diary content is extremely personal text data. Sending such intimate thoughts to external AI servers can create psychological resistance in users. Questions arise:

• What if my diary is read by others?
• Will my data be used to train AI?
• Can I trust this company with my deepest thoughts?

These concerns prevent many users from enjoying AI-powered journaling features.

Technical Solution

We deployed a lightweight Natural Language Processing (NLP) model client-side via WebAssembly. All text analysis logic operates entirely within the user’s device.

Technical Implementation:

// Dalendar emotion analysis pipeline
import { pipeline } from "@xenova/transformers";

// Load sentiment analysis model (runs in browser)
const classifier = await pipeline(
  "sentiment-analysis",
  "Xenova/distilbert-base-uncased-finetuned-sst-2-english",
);

async function analyzeEmotion(diaryText) {
  // Analyze text locally
  const result = await classifier(diaryText);

  // Map sentiment to emoji
  const emotion = mapToEmoji(result[0].label);

  return {
    emotion,
    confidence: result[0].score,
    timestamp: Date.now(),
  };
}

Model Details:

• Base Model: DistilBERT (compressed BERT)
• Size: 67MB (quantized from 268MB)
• Languages: Supports English, Korean, Japanese
• Accuracy: 91% emotion classification accuracy

User Value

Offline Functionality: You can write diary entries and receive emotion analysis even without internet connection. Perfect for private journaling anywhere, anytime.

Perfect Privacy: Your intimate thoughts never leave your device. No one—not even Realays—can read your diary. It’s “perfect privacy” guaranteed by technology, not just policy.

Insights Over Time: Track your emotional patterns:

• Mood trends over weeks/months
• Identify triggers for positive/negative emotions
• Visualize emotional journey
• Get personalized well-being tips

Technical Challenges & Solutions

Challenge 1: Model Size

• Problem: Full BERT model (400MB+) too large for web
• Solution: Used DistilBERT + quantization → 67MB
• Result: 6x smaller, 2x faster, 97% accuracy maintained

Challenge 2: Multi-language Support

• Problem: Need separate models for each language
• Solution: Multilingual DistilBERT with shared weights
• Result: Single 67MB model supports 3 languages

Challenge 3: Context Understanding

• Problem: Short diary entries lack context
• Solution: Implemented rolling context window
• Result: 15% accuracy improvement

Realays Vision: User-Centric AI

We pursue “User-Centric AI”, not technology-showing-off AI. This means:

1. Privacy First

Edge Computing:

• Process data on user’s device
• Never upload sensitive information
• Give users full control

Transparent Operations:

• Open about how AI works
• Allow users to inspect/delete data
• No hidden data collection

2. Accessibility for All

No Installation Required:

• Works in any modern browser
• No app download needed
• Cross-platform compatibility

Free Core Features:

• Basic AI features available to everyone
• No paywall for privacy
• Premium features enhance, not gate

3. Performance Excellence

Real-Time Response:

• Sub-100ms inference times
• Smooth 60 FPS interactions
• Progressive loading for large models

Optimized for All Devices:

• Works on 3-year-old phones
• Adaptive quality based on device
• Graceful degradation on older hardware

Technical Architecture Overview

FruitsFace Stack

User Photo → Browser
            ↓
      TensorFlow.js
            ↓
    Face Detection (MobileNet)
            ↓
  Landmark Extraction (68 points)
            ↓
   Fruit Matching Algorithm
            ↓
    Result Display

Dalendar Stack

Diary Text → Browser
           ↓
     WebAssembly Runtime
           ↓
   DistilBERT Model
           ↓
 Emotion Classification
           ↓
   Pattern Analysis
           ↓
  Mood Visualization

The Future of Privacy-Preserving AI

Edge Intelligence is the most powerful tool to realize user-centric AI, breaking away from:

• Privacy concerns of centralized AI
• High cost structures
• Dependence on internet connectivity
• Vendor lock-in

What’s Next for Realays

Short-term (2024):

• Voice emotion analysis in FruitsFace
• Advanced mood prediction in Dalendar
• Real-time collaboration features

Mid-term (2025):

• Federated learning across users
• Multi-modal AI (text + image + voice)
• Personalized AI assistants

Long-term (2026+):

• Fully autonomous edge AI
• Decentralized AI marketplace
• Open-source model contributions

Conclusion

Technology exists to serve people. At Realays, we’re committed to:

✅ Privacy - Your data stays yours ✅ Performance - Instant, delightful experiences
✅ Accessibility - AI for everyone, everywhere ✅ Innovation - Pushing web boundaries

We will continue to push the limits of web technology, bringing you faster, safer, and more enjoyable AI services.

Join us on this journey toward a future where AI serves users, not the other way around.

---

Want to try FruitsFace or Dalendar? Visit realays.com and experience privacy-first AI yourself!