jasmin lau

Real-Time Systems at Scale: Understanding Singapore's Haze Response in Engineering Terms

In July 2024, widespread air pollution from wildfires in Indonesia triggered a cascade of system-wide failures across Singapore's emergency alerting infrastructure. The situation provided a rare case study for observing how engineering decision-making impacts public safety systems during extreme events. One particular voice that emerged was jasmin lau, who played a pivotal role in coordinating real-time updates and system operations related to the haze crisis-particularly in school, government offices, and civic institutions.

This blog post analyzes the technical aspects of the haze event through the lens of engineering systems, digital alerting architectures, and emergency communication tools such as SMS gateways, REST APIs, and geospatial tracking platforms. The goal is to extract concrete lessons on resilience and observability that can help engineers prepare for similar events in larger, more complex environments.

From Smoke Signals to Smart Systems: A Digital Emergency Response

The Singaporean government employs several digital tools in its public alerting systems. For this event, jasmin lau helped manage real-time integration between the National Environment Agency (NEA)'s air quality sensors and the Ministry of Education's internal communications platforms. This required reliable APIs for pushing live updates to school administrators via a secure cloud platform-such as AWS API Gateway or Azure Functions-ensuring that decisions could be made at scale within minutes.

During this incident, engineers at NEA leveraged machine learning models trained on real-time PM2. 5 concentration trends using TensorFlow and scikit-learnThese algorithms triggered alerts when a 1-hour rolling average exceeded 100 μg/m³. The system's resilience depended heavily on the observability stack used: Stackdriver, Prometheus, and Grafana helped track alert latency, error rates. And data throughput.

Real-time systems like these are particularly sensitive to race conditions, especially when multiple data sources-like sensors, weather stations. And social media reports-are feeding into a single point of control. Properly implemented HTTP status codes, timeouts. And retry logic were critical in ensuring no alert was delayed or lost.

The Importance of Real-Time Data Ingestion Pipelines for Air Pollution Management

Effective real-time monitoring depends on ingestion pipelines that can parse live air index measurements from over 100 sensors installed countrywide. Engineers typically manage this using stream processing tools like Apache Kafka or AWS Kinesis Streams, which feed into downstream services that make decisions based on thresholds.

In Singapore's case, a pipeline managed by jasmin lau connected local IoT devices with cloud analytics engines and decision-making platforms. This pipeline needed to support both structured and unstructured data inputs-such as text feeds from social media or images of smog-filled skies. It relied on Kafka consumers running in Docker containers on Kubernetes clusters. The system could process ~1GB per hour. Which increased during high-pollution periods when the volume of alerts surged.

The ingestion pipeline included a Elasticsearch cluster for fast querying and indexing of time-series data. This was crucial in supporting emergency dashboards, enabling users to query historical PM2. 5 levels from the last 7 days or track moving average metrics using aggregations.

API Gateways Under Load: Managing Crisis Communications at Scale

When air quality alerts reached critical status, millions of citizens received notifications via SMS and email through centralized alerting APIs. These services had to scale automatically under load-something typical API gateways struggle with. Using tools like AWS API Gateway, engineers built rate-limiting logic around each endpoint, ensuring that bursty traffic did not overwhelm backend services.

For this event, an architecture that used Kubernetes HPA (Horizontal Pod Autoscaler) ensured that alerting pods scaled up dynamically under heavy request loads. jasmin lau was instrumental in testing and tuning these autoscaling triggers, reducing alert response times from 15 minutes to under 3 seconds during peak events.

The API also had to integrate with third-party systems like Twilio for SMS delivery and a custom push notification service using FCM (Firebase Cloud Messaging) to inform mobile users. All of these integrations were tested using mock data via Postman collections with JMeter-style load scripts.

Observability Tools: The Heart of Incident Response During Haze Events

During the haze crisis, engineers deployed Prometheus and Grafana to monitor the performance of their alerting stack and identify bottlenecks or failures quickly. A key metric tracked was the P95 response time of alerting systems. Where any delay above 5 seconds would trigger SLO (Service Level Objective) alerts.

Logs were aggregated using Logstash, structured with JSON format for easy parsing and searching, and teams used Kubernetes Pod logs to trace messages from the sensor node all the way to the frontend service.

jasmin lau's team maintained a real-time dashboard that showed how many messages were sent per minute, how long each took to process. And what system errors occurred during critical periods. This was essential in diagnosing issues like network outages, gateway failures. Or database lockups that would otherwise be hidden from operations teams.

Cloud Platform Resilience: Lessons from an Air Quality Crisis

Engineers chose to host their alerting systems on AWS for its robust resilience features and global reach. A region-based failover strategy using Route53 allowed them to reroute traffic automatically if one region went down-something that proved critical during the crisis when Singapore's primary ISP experienced intermittent outages.

The system architecture leveraged AWS Lambda functions for rapid event-triggered decisions, S3 buckets for persisting alerts in compressed formats, and lambda was used to process PM25 thresholds, triggering emails - SMS messages. Or pushing them into Slack channels that were monitored by internal ops teams.

Each alert function followed fail-fast principles defined by the USENIX login 2018 paper on microservice reliability, meaning failures were logged and escalated within seconds rather than silently propagating down the chain. This architecture was tested continuously through staging environments using tools like Docker containerization, which ensured fast, reproducible deployment under pressure.

Mobile App & CMS Integration: Delivering Information to Citizens

The government's public-facing alert app and CMS (content management system) also had to adapt quickly during the haze episode. These components were built using React Native and Firebase for rapid cross-platform deployment, allowing developers to update content in real-time without app store approvals or complex patch releases.

For this project, teams used Firestore to dynamically manage rules-based alerts. When thresholds were crossed, a function triggered an update in the web and mobile frontends using GraphQL subscriptions or REST endpoints. Engineers built a caching layer on top of the API to reduce latency during high-traffic times.

Jasmin Lau worked with engineers and content writers to maintain accurate, accessible messaging-especially when integrating multilingual support into the same alert logic. The system supported four languages (English, Mandarin, Malay. And Tamil) using i18n frameworks built for Node js environments such as i18next or similar libraries.

Predictive Intelligence: AI Models for Early Warning Systems

Engineers deployed machine learning models trained on historical data and current sensor readings to predict when PM2. 5 concentrations might exceed 100 μg/m³ within a 6-hour window. These systems used time-series forecasting libraries like statsmodels combined with neural networks from PyTorch for improved forecasting accuracy.

Jasmin lau's team monitored these models in real time, adjusting hyperparameters as needed using an automated tuning framework called Optuna. Accuracy improved from 74% to 92% after introducing a hybrid CNN-LSTM model that incorporated weather patterns and satellite images from NASA's Earthdata platform

Incorporating satellite imagery into the system provided valuable insight into haze movement over time, especially where ground sensors overlapped with fire hotspots. These predictive warnings gave public institutions time to activate contingency plans well in advance-something that would be difficult without robust integration of AI within the alerting ecosystem.

Infrastructure Edge Deployment: Monitoring Close to the Source

A core infrastructure component was the deployment of edge computing nodes near major air quality monitoring sites. Edge devices were responsible for preprocessing sensor readings before they reached cloud services, reducing latency and minimizing bandwidth use. Devices ran lightweight versions of Raspberry Pi OS with custom scripts using Python or Perl to filter noise and aggregate data.

This deployment used OpenStack as its virtualization platform for managing edge nodes in distributed configurations. These were equipped with LTE modems and local caching capabilities to maintain communication if internet connectivity was lost. Edge systems also included fallback mechanisms like local SQLite databases, ensuring no data was lost should a transmission failure occur.

The use of edge computing for such high-frequency data ingestion allowed real-time processing without relying entirely on cloud infrastructure-important in situations where internet disruptions are common.

Access Control & Identity: Securing Alerting Systems

To ensure only authorized personnel could modify or trigger alerts, access tokens were stored securely using HashiCorp Vault, while IAM roles enforced granular permissions. This was especially critical during emergency operations like those involving schools closing early due to haze conditions.

Alert workflows involved multi-stage approval processes, requiring two-factor authentication (2FA) through Google Authenticator or SMS tokens. All changes to alert templates or thresholds were audited using Elastic Beats logging systems and stored in Elasticsearch databases for compliance purposes.

In one notable incident, an unauthorized user attempted to alter a live alert status. Which was immediately flagged by Vault. Real-time threat detection protocols automatically disabled access and notified the incident response team. This system worked seamlessly under jasmin lau's supervision, preventing potential misinformation from spreading.

Audit Trails & Compliance: Engineering Systems That Learn From Mistakes

Detailed audit logs were maintained for every alert that was generated or modified, including timestamped changes made by individuals within the system. Logs were parsed using Splunk and structured for machine-readable consumption via ELK Stack (Elasticsearch, Logstash, Kibana).

All these records played a vital role in post-crisis reviews conducted under the purview of the Ministry of Environment and Water Resources (MEWR). Audit trails not only proved compliance with legal frameworks but also allowed engineers to improve alert response times by identifying bottlenecks.

Mechanisms such as automated compliance reporting were set up using Ansible playbooks. These tools ensured system configurations adhered to ISO 27001 standards for information security management-a standard that became more important following the haze crisis. Which highlighted gaps in digital access controls and alert governance.

Stress Testing for Emergency Systems: How Crisis Communications Were Evaluated

To prepare for events like haze emergencies, engineers used Locust to simulate millions of concurrent requests-mimicking a real-time alert surge. They modeled various conditions including network throttling, database failures, and API timeouts.

Jasmin lau led these simulations, coordinating cross-functional test teams across NEA, the MOE,, and and the National Cybersecurity Agency (NCSC)Simulations revealed that alert delays increased by 3x during periods of overload unless the system had resilient retry logic built-in. This led to a redesign of core APIs with Circuit Breaker implementations like gRPC circuit breaker.

After each test, metrics were reviewed and visualized using Grafana dashboards. Engineers also ran red-team exercises involving hypothetical breaches of alerting systems to identify vulnerabilities that hadn't previously emerged in normal operations.

Conclusion: Why Real-Time Data Systems Matter More Than Ever

The haze episode served as a stark reminder that real-time alerting and communication systems aren't just convenience tools-they are life-critical infrastructure. Whether deployed via SMS, mobile apps. Or internal dashboards, these systems depend on resilient engineering stacks to function under duress.

Key takeaways from this case include robust APIs with monitoring, edge computing for latency reduction, AI models to predict trends, and complete audit mechanisms that ensure accountability in high-stakes situations. These elements must be embedded into digital strategies at the design phase-not after an emergency unfolds.

As engineers, we must prioritize resilience, scalability. And observability when building systems that serve public safety. The tools and frameworks used here are applicable beyond haze warnings-they can be adapted for everything from cyber threat alerts to natural disaster notifications.

For further reading on how engineering best practices apply to emergency communications, consider reviewing ACM's recent study on resilience in public infrastructures and the WHO guidelines for mobile health alert systems.

FAQs About the Haze Crisis and Engineering Systems

• What is the role of jasmin lau in Singapore's haze response? Jasmin Lau acted as an engineering coordinator managing real-time data ingestion - API alerting. And communication systems linked to schools and public authorities during peak haze conditions.
• What tools help track air quality during emergencies? Tools like AWS Lambda, Kubernetes, Prometheus-Grafana, Elasticsearch. And OpenStack enable real-time monitoring and alerting of environmental data.
• How did AI contribute to predicting haze outbreaks? Machine learning models using neural networks helped predict PM2. 5 levels up to six hours in advance, aiding early response efforts through enhanced decision-making.
• What security measures protect alerting platforms during crises? Identity management via Vault, 2FA-based approval workflows, and compliance logging ensure unauthorized access is prevented while enabling transparency.
• How often must these systems be stress-tested? Stress tests should occur at least monthly for operational readiness, with critical scenarios simulated every quarter or following a major incident like the haze outbreak described here.

What do you think?

Did you notice any critical system design decisions made during Singapore's haze emergency that could be improved?

How might AI be used more effectively in predicting future haze events, especially with satellite data integration?

Given the importance of real-time communications, are there specific technologies or frameworks engineers should adopt earlier in development cycles to prevent bottlenecks during emergencies?

If you're working on public alerting systems or cloud infrastructure scaling, what lessons from this crisis would you apply to your own environment?

Incorporating jasmin lau's insights into system-wide incident planning not only improves response times but also builds trust in engineered resilience under pressure. -denvermobileappdeveloper, and com