This post was written with Felipe Monroy, Ana Jaime, and Nikita Gordeev from Snoonu.
Managing a massive product catalog in the ecommerce space has introduced new hurdles for retailers who are trying to efficiently connect customers with the items they truly want. Traditional one-size-fits-all approaches often result in lost opportunities and reduced customer engagement. For marketplace apps like Snoonu, personalization is crucial for driving customer engagement, improving conversion rates, and maximizing revenue per user while building lasting customer loyalty.
From the customer perspective, Snoonu’s users expect seamless product discovery experiences that save time and surface items that match their preferences. They want recommendations that are contextually appropriate and adapt to their changing needs. This makes personalization more than just a business imperative, but a customer satisfaction requirement.
In this post, we share how Snoonu, a leading ecommerce platform in the Middle East, transformed their product discovery experience using AI-powered personalization.
Challenges with ranking and recommendations
At Snoonu, innovation and excellence are guiding principles, as they strive to deliver the best shopping and delivery experience in Qatar. As they scale, optimizing ranking and recommending products as well as merchants to their customers is essential in enhancing user experience and driving business growth. Initially, Snoonu relied on basic rules like popularity-based ranking, which resulted in a recommendation system that lacked personalization and depth. To overcome this barrier, Snoonu turned to Amazon Personalize in 2024. They began with real-time recommendations for the entire platform by having a single, unified global model and progressively advanced toward recommendation models per vertical. This evolution showcases the agility and iterative approach of a startup mindset focused on continuously delivering incremental value. As a result, Snoonu achieved measurable impact through smarter, data-driven experiences that drastically boosted customer engagement and conversion rates.
The evolution of personalization at Snoonu
Snoonu began its personalization journey using static rules, primarily relying on popularity-based ranking. Although this method was straightforward to implement, it showed the same top products to all users, ignoring individual preferences, which lead to low engagement with recommended items. Perhaps most importantly, it failed to facilitate the discovery of long-tail products that could have been perfect matches for specific customers.
To address these limitations, Snoonu implemented real-time recommendations by developing a unified global model across all verticals (Marketplace, Food, and Groceries). The system generated daily recommendation lists for users. Although this approach improved relevance compared to static rules, it had its own challenges. Recommendations would become stale quickly, unable to capture nuanced user behavior or react to inventory changes. Furthermore, because recommendations remained unchanged throughout the week, the system’s inability to dynamically adapt to user actions and inventory updates limited its effectiveness in driving conversions.
The real breakthrough came with the decision to implement separate models for each vertical rather than maintaining a single, unified global model. This strategic shift acknowledged that customer behaviors and decision-making processes vary significantly across verticals—for instance, a customer ordering lunch exhibits different patterns compared to someone planning weekly groceries. By developing specialized models for food delivery, grocery shopping, and marketplace purchases, Snoonu could better capture the unique nuances of each vertical’s customer journey, despite some overlapping behaviors. This vertical-based approach produced more precise recommendations, ultimately enhancing customer experience and satisfaction. Additionally, Snoonu transformed their training strategy from a 3-week cycle to daily updates—a crucial change that we explore later in this post.
The system was further enhanced with sophisticated filter expressions to improve recommendation relevance. One key implementation of this feature is used to exclude items that were was already added to a basket to reduce the exploration list. Additionally, the filter expressions support recommendations on the user’s current browsing category. For example, when a user searches for laptops, the system intelligently suggests related electronics items.
To optimize performance and scalability, Snoonu implemented Redis caching through Amazon ElastiCache. This addition significantly reduced API latency, improved system scalability, and optimized costs while enhancing the overall user experience.
Solution overview
Snoonu’s architecture uses Amazon Personalize to deliver real-time recommendations by seamlessly capturing user interactions, continuously updating its personalization models, and filtering by available products to its customers. The following diagram illustrates the solution architecture.
The technical solution architecture comprises several key stages using AWS services and external tools: data preparation and collection, model training, and real-time recommendations. In the following sections, we discuss these stages in more detail.
Data preparation and collection
Snoonu implemented a dual-pipeline data collection strategy for their recommendation system. Amplitude captures all user interactions across the app, which feed into Snoonu’s primary pipeline that exports daily data to BigQuery for historical storage and model retraining. The historical storage comprises structured datasets of user-item interactions, items metadata, and users metadata, which are processed daily in Databricks. After validation and feature transformation, the data is exported to Amazon Simple Storage Service (Amazon S3) and used to retrain the model daily through Amazon Personalize. For more details about the different data types used and how they power personalization, see Preparing training data for Amazon Personalize.
Daily dataset updates, collected by this data preparation and collection pipeline, support fresh item catalogs and updated business rules.
Model training and update
Snoonu implemented multiple Amazon Personalize recipes for different use cases. The aws-user-personalization recipe powers homepage ranking and cart and product detail page suggestions. The aws-similar-items recipe handles “frequently bought together” recommendations, and the aws-personalized-ranking recipe enhances category and subcategory pages ranking.
Although Amazon Personalize typically recommends weekly training schedules, through extensive experimentation, Snoonu discovered their use case required frequent updates. They transitioned from a 3-week training cycle to weekly, and finally settled on daily training to maintain relevancy across all verticals.
Each vertical uses different Amazon Personalize recipes optimized for their specific needs. For instance, Snoonu’s Marketplace implements both user personalization and reranking capabilities to handle the extensive catalog effectively. Their grocery vertical primarily uses similar items that recommends “frequently bought together” items to enhance basket building, and the food delivery service relies on user personalization to capture ordering patterns. To learn more about choosing the right Amazon Personalize recipe, see Choosing a recipe.
To complement this and update the real-time recommendation system, they stream Amplitude events that are filtered by business verticals through Amazon Kinesis Data Streams, which then flow into AWS Lambda. The Lambda function performs three critical tasks: it validates the schema to make sure incoming Kinesis Data Streams data is well-formed, processes these events before sending to Amazon Personalize using the PutEvents API to update the event tracker, and triggers recommendation updates by sending HTTP POST requests with userIds to their intermediary service. This enables the system to continuously update and refine recommendations based on immediate user behavior.
Real-time recommendations
Given that Snoonu is using a custom recipe, they built a campaign that hosts a solution version to return real-time recommendations. To learn more about this process, see Custom resources for training and deploying Amazon Personalize models.
To effectively serve these recommendations at scale, Snoonu developed an intermediary service that acts as a smart bridge between the frontend and their campaign solution. This service addresses the challenges of instantly responding to user behavior while efficiently managing caching, API usage, and response times at scale. To achieve this, they use Amazon ElastiCache for caching recommendations, implementing a live trigger cache invalidation strategy instead of traditional Time to Live (TTL) logic. When a user interacts, a Lambda function triggers the recommendation service, potentially clearing the cache and fetching new recommendations only if the action is likely to change the user’s preferences.
The service also handles postprocessing of Amazon Personalize results, filtering by factors not natively supported such as geographical availability and real-time stock status. This architecture makes sure the system can instantly respond to user behavior while efficiently managing caching, API usage, and response times at scale, providing personalized recommendations that stay fresh and relevant to users.
Finally, all recommendation requests are logged through Amazon MQ (using RabbitMQ), so Snoonu can compare pre-and post-filtering recommendations and monitor system performance.
Business outcomes
The implementation of Amazon Personalize delivered remarkable business results for Snoonu, demonstrating the powerful impact of personalized recommendations in ecommerce. In the Groceries vertical, the platform witnessed a striking 1,600% increase in add-to-cart events from cart recommendations, indicating substantially improved customer engagement and product discovery. The personalization strategy proved to be highly cost-effective, generating a Gross Merchandise Value (GMV) that was 47 times the total model investment over a 6-month period. This translated to QR 2.6 million (USD $715,000) in accumulated incremental GMV directly attributed to recommendation-driven conversions. Most notably, personalized recommendations contributed 30% to customer basket size in orders with at least one recommended product. These results also showcase how effectively personalization can create a multiplier effect, driving immediate sales and encouraging customers to explore broader product categories and make more informed purchasing decisions. The success in the marketplace paved the way for Snoonu to expand this personalization strategy across other verticals, using customer insights to create more engaging and relevant shopping experiences.
Key learnings
The journey to implementing AI-driven personalization yielded valuable insights. Starting small and scaling fast proved crucial, helping Snoonu validate impact before expanding to more complex use cases. Data quality emerged as a critical success factor, requiring significant investment in preparation and maintenance of consistent schemas.
Snoonu’s journey with personalization began with a single, unified model across all verticals, which provided a foundation for understanding customer behavior patterns. As the platform matured, the team strategically evolved their approach by developing specialized models for different verticals and implementing specific recommendation recipes tailored to various use cases.
Conclusion
Snoonu’s journey with Amazon Personalize demonstrates the transformative power of AI-driven personalization in ecommerce. Snoonu’s phased approach to personalization substantially impacted customer behavior. This solution optimized performance within specific verticals and created a multiplier effect, enhancing customer engagement and encouraging cross-category exploration. Amazon Personalize helped Snoonu achieve substantial business value while significantly improving customer satisfaction. This success story opened the door to explore even more innovative AI-powered capabilities, such as using generative AI models on Amazon Bedrock, to drive further enhancements to their personalization and discovery experiences.
Learn more about how Snoonu used Amazon Bedrock to categorize over 1 million products.
About Snoonu
Snoonu is Qatar’s first All-in-One delivery app. Since their founding in 2019 by Mr. Hamad Al-Hajri, Snoonu has evolved into the country’s leading tech-driven super-app, offering a wide range of delivery services—including food, groceries, electronics, pharmaceuticals, and more.
About the Authors
Felipe Monroy is a Senior Data Scientist at Snoonu. He has extensive experience in data science and ML engineering, holding diverse roles at global companies such as McKinsey & Company and Rappi across Latin America and Australia. With a master’s degree in Data Science and Innovation from the University of Technology Sydney. He specializes in advanced analytics, leveraging expertise in MLflow, AWS, and Databricks. He is particularly passionate about MLOps and LLMs, frequently contributing to discussions and mentoring others in these areas.
Ana Jaime is the Head of AI & Data Science at Snoonu, with a strong foundation in both technical execution and strategic leadership. Ana held key leadership roles at Rappi before joining Snoonu. She is now dedicated to building and leading a high-performing AI & Data Science team, with projects spanning personalization, ETAs, demand forecasting, content moderation, and item categorization. Through her leadership and the team’s contributions, Snoonu was recognized as the Best Emerging Tech Company in AI in 2024. Ana is committed to fostering talent with a keen interest in ground-breaking technologies, passion, and dedication.
Nikita Gordeev is the Chief Technology Officer at Snoonu. He has over 10 years of experience in software development and leadership roles in the banking, telecommunication, and ecommerce sectors. He holds a master’s degree in Information Security and Advanced Studies at MIT. He is always open and eager to share knowledge and insights with those who are interested.
Saubia Khan is a Startup Solutions Architect residing in the sunny city of Dubai. Her focus is AI, ML, and generative AI, where she is passionate about making it easier for her customers to build and deploy their AI solutions on AWS.
Ahmed Azzam is a Senior Solutions Architect at AWS. He is passionate about helping startups not only architect and develop scalable applications, but also think big on innovative solutions using AWS services.
