Machine Learning at Hike: Five AI Systems Serving Millions

As the leader of the Machine Learning team at Hike Limited, I spearheaded the development of five major AI-driven features that transformed user experience across the platform. These projects spanned computer vision, natural language processing, social matching, trust & safety, and gaming - collectively serving millions of users across India.

1. Hikemoji: Computer Vision for Avatar Generation #

Project Overview #

Hikemoji generated personalized avatars directly from users’ selfies. My role focused on developing sophisticated computer vision models to match avatar components to specific facial attributes.

###Technical Approach

Core Technologies:

• Python, TensorFlow, PyTorch, OpenCV
• BigQuery for data storage
• Airflow for workflow orchestration

Key Components:

1. Facial Feature Extraction: Models to identify and map key facial features from selfies
2. Component Matching Algorithm: AI-driven system matching facial features with avatar components
3. Style Transfer Techniques: Algorithms adapting avatar aesthetics to user preferences
4. Real-time Processing: Optimized models for quick, on-device generation

Challenges and Solutions #

• Challenge: Accurate facial detection across diverse demographics

  • Solution: Trained on diverse datasets with data augmentation techniques

• Challenge: Balancing accuracy with artistic appeal

  • Solution: Developed scoring system balancing facial similarity with aesthetics

• Challenge: Mobile device performance

  • Solution: Model compression and TensorFlow Lite optimization

Results #

• 95% user satisfaction rate
• 70% increase in avatar feature engagement
• Avatar creation time reduced from minutes to seconds
• 1M+ unique avatars processed in first month

2. Vernacular Sticker Keyboard: NLP and Federated Learning #

Project Overview #

Developed an AI-driven vernacular sticker keyboard that intelligently suggested stickers based on multilingual inputs, including Hinglish, Tamil English, and various language combinations.

Technical Approach #

Core Technologies:

• Python, TensorFlow, TensorFlow Lite
• NLP techniques for language understanding
• Federated learning for privacy-preserving updates

Key Features:

1. Multilingual Input Processing: NLP models understanding mixed-language inputs
2. Contextual Sticker Suggestion: AI model suggesting relevant stickers based on text and context
3. On-Device Personalization: TensorFlow Lite models for on-device learning
4. Federated Learning: System for updating global models while maintaining privacy

Challenges and Solutions #

• Challenge: Handling diverse linguistic combinations

  • Solution: Trained on vast multilingual corpus with advanced tokenization

• Challenge: Real-time mobile performance

  • Solution: TensorFlow Lite optimization and efficient caching

• Challenge: Balancing personalization with privacy

  • Solution: Federated learning allowing improvements without centralized data

Results #

• 40% increase in sticker usage platform-wide
• 60% improvement in suggestion relevance
• Successfully handled 10+ language combinations
• Privacy maintained through federated learning

3. Vibe Metaverse: Social Matchmaking #

Project Overview #

Developed a sophisticated AI-driven matchmaking system for Vibe, Hike’s metaverse friendship network. The goal was creating meaningful connections by optimally selecting users for virtual rooms based on interests, interaction history, and social dynamics.

Technical Approach #

Core Technologies:

• Python, optimization solvers
• BigQuery, Airflow, TensorFlow

Key Components:

1. User Profiling: Comprehensive profiles based on interactions, preferences, and behavior
2. Matchmaking Algorithm: Advanced optimization algorithm for optimal room grouping
3. Real-time Processing: Real-time matchmaking decisions
4. Performance Metrics: KPIs measuring match success and user satisfaction

Challenges and Solutions #

• Challenge: Balancing multiple matchmaking factors

  • Solution: Multi-objective optimization with weighted importance

• Challenge: Ensuring diversity while maintaining relevance

  • Solution: Constraint-based approach mixing similar and diverse users

• Challenge: Dynamic user preferences

  • Solution: Adaptive system continuously updating profiles

Results #

• 50% increase in virtual room engagement
• 40% improvement in social interaction satisfaction scores
• 85% average room satisfaction rate
• 60% reduction in inactive/abandoned rooms

4. Vibe Trust & Safety: Malicious Reporting Detection #

Project Overview #

Developed a sophisticated AI system to detect and mitigate malicious reporting within the Vibe metaverse, maintaining a safe, trustworthy environment.

Technical Approach #

Core Technologies:

• Python, modified PageRank algorithm
• BigQuery, Airflow, TensorFlow

Key Components:

1. Trust Scoring System: Modified PageRank assigning trust scores based on interactions and reporting history
2. Behavioral Analysis: Models analyzing user behavior patterns and identifying anomalies
3. Report Classification: ML model classifying reports by genuineness likelihood
4. Real-time Processing: Real-time analysis and decision-making

Challenges and Solutions #

• Challenge: Distinguishing genuine from false reports

  • Solution: Multi-faceted approach combining trust scores, behavioral analysis, content evaluation

• Challenge: Evolving malicious behavior

  • Solution: Adaptive system continually updating through machine learning

• Challenge: Balancing swift action against false positives

  • Solution: Tiered response system with human oversight for high-stakes decisions

Results #

• 75% reduction in false/malicious reports (first 3 months)
• 40% improvement in user trust scores
• 60% faster resolution of legitimate reports
• 99.9% accuracy distinguishing genuine vs. malicious reports

5. Rush Gaming: AI-Driven Player Matchmaking #

Project Overview #

Developed an innovative AI-driven matchmaking system for Rush, Hike’s real-money gaming network. The goal was creating fair, engaging, personalized gaming experiences by matching players based on skill levels, gaming behavior, and overall experience.

Technical Approach #

Core Technologies:

• Python, TensorFlow
• BigQuery, Airflow
• Custom ranking algorithms (ELO, TrueSkill-inspired)

Key Components:

1. Player Skill Evaluation: Multi-faceted rating system considering game-specific skills and performance
2. Behavioral Analysis: Models analyzing play style, preferences, interaction patterns
3. Real-time Matchmaking Engine: High-performance system for instant decisions
4. Fairness Assurance: Algorithms ensuring balanced matches, detecting unfair advantages
5. Adaptive Learning: Continuous learning from match outcomes and feedback

Challenges and Solutions #

• Challenge: Balancing match quality with wait times

  • Solution: Dynamic algorithm adjusting criteria based on queue times and player pool

• Challenge: Ensuring fairness in diverse player ecosystem

  • Solution: Multi-dimensional ranking beyond win/loss ratios

• Challenge: New player onboarding

  • Solution: Rapid assessment using initial games to gauge skill levels

Results #

• 40% increase in player retention
• 60% improvement in match quality ratings
• 30% reduction in queue times
• 50% reduction in negative gaming experiences

Common Technical Infrastructure #

All five systems leveraged shared infrastructure:

• BigQuery: Large-scale data storage and analysis
• Airflow: Workflow orchestration and scheduling
• TensorFlow/PyTorch: Model development and training
• Python: Core development language

Conclusion #

These five ML systems demonstrate the breadth and impact of AI in transforming mobile platform experiences. From computer vision and NLP to social matching, trust & safety, and gaming - each project solved unique challenges while serving millions of users.

The success of these initiatives at Hike showcases how thoughtful application of machine learning can enhance user engagement, ensure platform integrity, and create more personalized, meaningful experiences at scale.

About the author: Dipankar Sarkar is a technology leader with deep expertise in machine learning and AI. As ML team leader at Hike, he pioneered multiple AI-driven features including matchmaking, computer vision, and NLP systems serving millions of users. View all posts | Get in touch