Towards delivering world-class recommendations at the lowest possible cost

A Weekly Tech Blog Series on Innovation and Cost Efficiency

Sharechat is India’s largest homegrown social media company with native and regional user generated image and video content. Our two apps, ShareChat and Moj, are used by over 325 million users in 15 regional languages across India. We have millions of posts a month created on our platforms and it is critical that we serve the most relevant, personalized content to our users in order for them  to derive value from our platforms, and to drive long-term revenue growth for the company via ads and in-app transactions.

Delivering high quality personalized recommendations to such a large user base is only possible via training and deploying advanced AI models at scale that can learn and adapt to shifting user behaviours and preferences so that we show the most relevant and engaging content for a given user at any given time. This requires building realtime and batch model training pipelines that consume and process large amounts of features and user interaction event streams, and serving those models in production in a multi-tiered fashion to generate personalized feeds. Developing and running such a system entails a large amount of infrastructure cost across compute, data storage and processing, networking and licensing.

In 2024, ShareChat’s primary focus was getting as close as possible to operational profitability and driving sustainable growth from there. In order to achieve this as quickly as possible, we pushed towards reducing our server costs while continuing to grow our revenue at the same time (which also implied growing user retention, timespent and engagement driven by feed). Feed recommendations are the backbone of our apps but also the most expensive system in terms of our server cost footprint and hence we asked the team to focus on the problem of delivering the best personalized recommendations to our users in the most efficient manner possible.

What the team managed to deliver through the year went beyond the most aggressive goals we had set for the team at the start of the year. In 2024, we achieved:

• 76% reduction in ShareChat feed server costs while causal absolute D1 user retention lift from personalized recommendations increased to 1.7x of what it was at the start of the year
• 80% reduction in Moj feed server costs while causal absolute D1 user retention lift from personalized recommendations increased to 1.2x of what it was at the start of the year
• We also reduced a lot of complexity in our systems which gives us a much better base to iterate and build upon for the future. Albeit hard to quantify, this is a huge benefit by itself.

This was a long journey with many technical, execution and organizational challenges and obstacles, but this took us a long way towards making the ShareChat app EBITDA profitable and bringing the company close to full operational profitability as outlined in our recent announcement. This puts our business on a much stronger footing to focus on building for growth in 2025 and beyond while maintaining overall profitability and sustainability.

On the technical and execution front, there was no one silver bullet but rather many improvements that added up to give us the impact we needed – as is often the case with recommendation systems in general.  Some of these were:

• A large number of system optimizations across the entire rec sys stack (i.e. ranker training and inference, delivery systems, feature computation and serving, retrieval, etc) including ablations and deprecations of low value components
• Adoption and migration to various open source solutions as well as BYOC deployments in order to reduce managed services and licensing costs
• Improved cost observability and attribution which provided visibility into the highest impact areas to optimize, and reorganizing our teams with narrowed focus on efficiency of those components
• Establishing a framework for long-term revenue vs. cost tradeoffs of all proposed changes with a heavy reliance on A/B testing and causal data
• Unification of many systems and codebases across apps and teams to reduce complexity and maintenance burden thus increasing execution speed
  

In this series of technical blogs, we will delve deeper into some of the major optimizations we made to our recommendation systems including ideas that are specific to how cloud costs are billed and tracked. We expect this will be useful for anyone looking to build high quality and efficient recommendation systems and reduce their infrastructure costs. Stay tuned!

We are hiring! Check out open roles at https://www.linkedin.com/company/sharechat/jobs