Scaling a super app with AI at the core of the experience

Snoonu is building a super app designed to simplify everyday life across the Gulf Cooperation Council (GCC). By bringing food, groceries, retail, and services into a single platform, Snoonu serves hundreds of thousands of customers across Qatar while expanding into Kuwait, Oman, and Bahrain. At the same time, the company helps more than 8,000 local and home-based businesses grow digitally and provides income opportunities for over 7,000 drivers and their families.

As Snoonu continues to grow across the region, its Data Science and AI team plays a central role in turning data into a competitive advantage across personalization, customer experience, monetization, and operational efficiency. AI is deeply embedded across the platform. Genie, Snoonu’s AI shopping assistant, helps customers discover products through text and image-based interactions, while Smart Catalog agents continuously improve product quality through categorization, attribute extraction, and content moderation in the background. These systems run at scale in production on Amazon Web Services (AWS), directly shaping real user experiences. “Our goal is to make Snoonu the most intelligent and reliable super app in the region, and AI is at the center of how we are getting there,” Ana Jaime, Head of AI & Data Science at Snoonu, explains.

“Our goal is to make Snoonu the most intelligent and reliable super app in the region, and AI is at the center of how we are getting there.”

Limited visibility slowed innovation at scale

As these AI agents and ML systems became more sophisticated and central to the product experience, Snoonu needed a clearer understanding of how they behaved in real-world environments. However, visibility gaps across both domains made this difficult.

For AI agents, the team lacked insight into how conversations unfolded in production, including how models reasoned and how tool calls were executed. When issues occurred, engineers had to manually replicate user sessions and analyze behavior locally, which made debugging slow and resource intensive. “When a problem in production was caused by a user, it was very difficult to know what the agent was doing,” says Felipe Monroy, Senior Data Scientist at Snoonu.

Latency spikes were especially difficult to diagnose. In some cases, responses took up to a minute, but the team could not easily determine whether the issue came from the model, orchestration logic, or integrations.

At the same time, ML systems such as estimated time of arrival (ETA) prediction models faced a different challenge. These models help determine when deliveries will arrive and directly influence customer expectations, yet the team lacked real-time visibility into prediction behavior. Engineers could not easily monitor which ETA values were being returned in production or validate model performance after updates. Instead, they relied on manual warehouse queries with delays of up to a day. “We previously had to run manual queries and wait to understand what was happening in production,” adds Juan David Herrera Parra, Senior ML Backend Engineer at Snoonu.

This slowed feedback loops, made it harder to validate changes, and limited the team’s ability to quickly detect and resolve issues.

“We previously had to run manual queries and wait to understand what was happening in production.”

Real-time visibility across AI agents and ML systems

With Datadog, Snoonu unified observability across its entire platform, gaining real-time visibility into both AI agents and ML systems within a single environment.

Datadog Agent Observability gives the team end-to-end insight into how Genie behaves in production. Engineers can trace interactions across prompts, model outputs, and tool calls, while using metadata and traces to quickly investigate issues and identify where failures occur. “It is very straightforward to search the span and understand what happened across the interaction flow,” says Monroy.

This level of visibility allows Snoonu to understand agent behavior in full context and quickly identify the root causes of issues. When analyzing long-running sessions, the team discovered patterns tied to extended conversations and unexpected user inputs. Using these insights, they refined prompts and introduced summarization strategies to improve response times and overall efficiency. Agent Observability also improved testing and deployment workflows, enabling teams to validate behavior in staging and confirm performance in production with greater confidence. “Before, we were only confident checking locally, but now we can test everything in staging and see all the traces in one place,” explains Monroy.

In parallel, Datadog Custom Metrics provides real-time visibility into Snoonu’s ML systems. Engineers can now monitor prediction latency, track model outputs, and observe service behavior as it happens. This is especially critical for ETA predictions, where even small changes in model output can directly impact customer experience. “They give us a real-time view of the full service landscape so we can make adjustments on the fly,” says Parra.

With this visibility, the team can validate daily model retraining, detect unexpected shifts in predictions, and respond quickly to changing conditions. Custom Metrics also enables proactive issue detection. In one case, Snoonu surfaced an unusual spike in memory consumption during a routine monitoring review. By tracking the pattern across multiple deploys and correlating it with service-level metrics, the team confirmed a memory leak in a critical service and resolved it before it impacted service reliability and customer experience. By combining Custom Metrics with APM, logs, dashboards, and alerts, Snoonu can move seamlessly from high-level monitoring to deep root cause analysis across all services.

Faster iteration, stronger performance, and confident growth

With full visibility into both AI agents and ML systems, Snoonu has transformed how it builds and operates its platform. “We can now trace exactly what agents are doing at every step, including all tool calls, so that we can quickly identify where issues happen,” adds Monroy.

“We can now trace exactly what agents are doing at every step, including all tool calls, so that we can quickly identify where issues happen.”

This has reduced debugging time for agent-related issues by up to 70% and improved collaboration between QA and backend teams. At the same time, real-time monitoring of ML systems has enabled the team to identify bottlenecks and optimize performance. For the ETA ML model, Snoonu reduced latency by up to 68%, improving responsiveness across key user experiences.

More broadly, Snoonu has adopted a more iterative and proactive approach to development. Teams can release features earlier, monitor their impact in production, and continuously improve based on live data rather than delayed insights. This has increased confidence across the organization and reduced the risk associated with scaling AI systems. “The biggest change is the level of confidence we have when shipping code to production,” adds Parra.

As Snoonu continues to expand across the GCC and beyond, this unified observability approach allows the team to innovate faster while maintaining a high standard of quality. “Now that we have full visibility into how our AI agents and ML systems perform, we can continuously improve and expand what’s possible for our users. This is only the start,” concludes Jaime.

“Now that we have full visibility into how our AI agents and ML systems perform, we can continuously improve and expand what’s possible for our users. This is only the start.”

With contributions from Ana Jaime, Felipe Monroy, Juan David Herrera Parra, and Nikita Gordeev