Forward Systems

Recommender System for a 
300-Million-User Social Platform

We have built a scalable recommender system for a social platform, utilizing user interaction data to deliver personalized content suggestions and drive higher engagement.

Python

BERT

FastAPI

MySQL

Google Cloud

Problem

Ask.fm is a popular social networking platform where millions of users ask and answer questions daily. 

As more and more of users were asking and answering questions daily, content discovery was becoming increasingly challenging. 

The platform needed a way to identify and group semantically similar questions across different languages all while handling massive amounts of data in real-time.

solution

We developed a recommender system that improved semantic matching across multiple languages, significantly boosting user retention and interaction on the platform while maintaining scalability and performance during high traffic periods.

300 million active users

150 countries

49 languages

Our Approach

  • Machine Learning for Text Similarity: We implemented BERT-based models to map questions into vector spaces, enabling high-precision detection of semantically similar questions. Cosine similarity scoring was used to compare these vectors and match similar content.
  • Multilingual Support: Using multilingual embeddings, our solution supported semantic matching across different languages, allowing Ask.fm’s diverse user base to benefit from enhanced content discovery, regardless of their language preferences.
  • Performance Optimization: To ensure real-time processing at scale, we:
  • Selected efficient models to reduce dimensionality and computation time.
  • Implemented vector caching and indexing to minimize repetitive calculations.
  • Pre-filtered data by age or category, speeding up query performance.
  • Optimized SQL queries for batch data retrieval.
  • Grouped vector search by shard for parallel processing, allowing horizontal scaling.
  • Applied load balancing to ensure system stability during high-traffic periods.
  • The system was built with Python and FastAPI and deployed on Google Cloud Run. We used Google Cloud Build for the CI/CD pipeline, ensuring seamless integration with Ask.fm’s backend infrastructure.
  • We used Scrum software development framework to achieve fast delivery and address any issues promptly.
  • We ensured that all data processed by the system adhered to strict data protection regulations through anonymizing and securely handling the user data.

Ask.fm, a global social networking platform, with a user base of over 300 million active members spanning more than 150 countries and supporting 49 languages.

"

I highly recommend this company to any business looking for a machine learning integration. Their ability to tackle any problem is remarkable. It helped our team to come up with the most efficient solution for question distribution.

TOMAS JUNDO, HEAD OF PRODUCT AT ASK.FM

Let’s discuss your bespoke software requirements...

Any sector, any size, any software development challenge.