What Our Clients Say

challenge

A KL-based fintech was running fraud detection models on a single GPU instance. Peak transaction volumes caused model inference delays of over 8 seconds, triggering false declines and customer complaints.

solution

nexurars designed a multi-node inference architecture with load balancing and auto-scaling based on queue depth. Infrastructure was codified in Terraform with a staging environment that mirrored production.

results

Inference latency dropped to under 200ms at peak. False decline rate fell by 62%. The internal team now manages scaling independently using the runbooks provided. Engagement took 10 weeks.

challenge

A Penang-based manufacturer had a working predictive maintenance model in a Jupyter notebook but no way to deploy, retrain, or monitor it in production. The data scientist who built it had moved on to a new role.

solution

We built an MLOps pipeline that automated data ingestion from factory sensors, scheduled retraining, tracked experiments, and deployed updated models through a CI/CD flow. All config was version-controlled.

results

Model retraining went from a manual weekly process to an automated daily cycle. Unplanned downtime dropped by 34% in the first quarter. The operations team can now trigger retraining themselves.

challenge

An e-commerce company in Kuala Lumpur was spending RM 38,000 per month on cloud resources for their recommendation engine and search ranking models, with no clear understanding of cost drivers.

solution

nexurars performed a two-week cost audit, identified idle training instances running 24/7, oversized inference nodes, and missed spot instance opportunities. Delivered a prioritized action plan with effort estimates.

results

Monthly cloud spend dropped to RM 22,400 — a 41% reduction. The largest single saving came from scheduling training instances to shut down overnight, which alone saved RM 7,200/month.