foodpanda has deployed a data- and AI-driven rider safety system across its Asia-Pacific operations and delivery partners across 10 markets to improve on-road safety outcomes.
The initiative supports delivery partners across 10 markets and is designed to shift rider safety from reactive incident management to predictive risk prevention.
foodpanda APAC’s senior director of logistics, Anson Chin, told iTnews Asia that the company leverages telematics data in Singapore and Malaysia, and app usage data, including trip acceptance, idle times, and route choices, to identify unusual patterns or situations that may require attention.
In some markets, the company uses motion-based signals to enhance risk detection.
According to Chin, the company has seen a 30 percent reduction in delivery partner accidents across APAC.
It has seen rider safety satisfaction increase from 46.4 percent to 51.7 percent in Singapore.
Data architecture supporting rider safety
Each market processes its own data through a localised ingestion pipeline via secure SDKs and encrypted APIs embedded in the rider app.
This ensures data is collected responsibly and in accordance with each country’s regulatory requirements.
“Once ingested, raw operational and behavioural data is stored in secure cloud environments,” said Chin.
“From there, it is structured and processed within a centralised data lake, with region-specific partitions to meet GDPR, PDPA, and other regulatory obligations,” he added.
To support real-time risk monitoring, the team use stream processing technologies to detect anomalies or potential safety risks as they happen. This is complemented by batch processing pipelines, which train AI models and enable deeper retrospective analysis to continually refine our safety approach, said Chin.
The architecture helps shift from reactive safety measures to a proactive, predictive system.
Training of models
The company use AI models trained on historical riding patterns to help identify behaviours which may indicate risk.
These models look at changes in speed or movement - such as sudden accelerations, harsh braking, or consistently riding above typical speed ranges for the area, explained Chin.
Instead of looking at just single actions, the model builds a broader view of the trip and assigns a rider safety score that updates continuously.
From the rider’s point of view, the use of AI is designed to feel supportive rather than supervisory. While the model continuously assesses trip-level riding patterns, the insights are used to reduce exposure to higher-risk situations, including reallocating demanding routes or loads, and to offer timely safety nudges when conditions suggest increased risk.
- Anson Chin, Senior Director of Logistics, APAC, foodpanda
The data is not used for punishment or retroactive enforcement, but as a tool to help riders stay safe and confident on the road, Chin said.
The company also determine what counts as “safe” versus “risky”.
The team calibrates thresholds using a mix of statistical analysis that identify outliers and anomalies against market-specific driving norms, as riding behaviour can differ across countries, and verified accident and incident reports to identify recurring high-risk patterns.
Chin said that this helps ensure that the model reflects real, on-ground behaviour instead of relying on one global standard for all markets.
“Our algorithms also factor in real-time rider position and speed, along with a proprietary model that estimates remaining drive time based on vehicle type, time of day/day of week, and the specific leg of the journey. This helps estimate how much time the rider will require to drive to the next point, to avoid assigning orders in a way that encourages speeding or unsafe rushing,” Chin added.
The goal is to ensure riders have enough time to travel safely without pressure.
The company also considers physical capacity limits.
For example, if an order is too heavy or bulky for a rider’s vehicle type, the system automatically avoids assigning it to them or splits the delivery assignment to an acceptable share.
This prevents situations where riders might struggle with loads that aren’t safe to carry.
Hybrid infrastructure model
Foodpanda uses a hybrid infrastructure model that enables a centralised global operation whilst allowing adjustments to meet each country’s regulatory and operational needs.
“We maintain a centralised data and AI layer which includes data infrastructure, AI models and global governance policies managed by our global data science team,” said Chin.
“The markets then run a federated localisation layer with configurable compliance layers that can adapt the system to local rules and performance requirements. These include country-specific data residency, storage duration, and model retraining cadence,” he added.
All markets connect through a shared integration architecture using APIs and local dashboards for safety reporting and regional ops monitoring.
This hybrid setup provides global capabilities while maintaining flexibility for compliance and performance in every market.
Looking ahead, the company’s approach is to stay adaptable and test what works in each market, learning from real-world conditions, and gradually integrating AI enhancements where they add clear value.
Chin said, the diversity of markets the company operates in means there can be no one-size-fits-all solution, so the long-term direction is about building systems that are flexible, locally relevant, and capable of evolving as technology and rider needs continue to change.
Chin added that the next phase for foodpanda is not about adopting more AI for its own sake, but about applying it thoughtfully to create safer, supportive working environments across the entire gig ecosystem.
